Encoder-free multimodal large language models (MLLMs) eliminate the need for a well-trained vision encoder by directly processing image tokens before the language model. While this approach reduces computational overhead and model complexity, it often requires large amounts of training data to effectively capture the visual knowledge typically encoded by vision models like CLIP. The absence of a vision encoder implies that the model is likely to rely on substantial data to learn the necessary visual-semantic alignments. In this work, we present BREEN, a data-efficient encoder-free multimodal architecture that mitigates this issue. BREEN leverages a learnable query and image experts to achieve comparable performance with significantly less training data. The learnable query, positioned between image and text tokens, is supervised by the output of a pretrained CLIP model to distill visual knowledge, bridging the gap between visual and textual modalities. Additionally, the image expert processes image tokens and learnable queries independently, improving efficiency and reducing interference with the LLM’s textual capabilities. BREEN achieves comparable performance to prior encoder-free state-of-the-art models like Mono-InternVL, using only 13 million text-image pairs in training—about one percent of the data required by existing methods. Our work highlights a promising direction for data-efficient encoder-free multimodal learning, offering an alternative to traditional encoder-based approaches.

Recent progress in vision Transformers exhibits great success in various tasks driven by the new spatial modeling mechanism based on dot-product self-attention. In this paper, we show that the key ingredients behind the vision Transformers, namely input-adaptive, long-range and high-order spatial interactions, can also be efficiently implemented with a convolution-based framework. We present the Recursive Gated Convolution (${\mathit{g}}^{\mathit{n}}$gnConv) that performs high-order spatial interactions with gated convolutions and recursive designs. The new operation is highly flexible and customizable, which is compatible with various variants of convolution and extends the two-order interactions in self-attention to arbitrary orders without introducing significant extra computation. ${\mathit{g}}^{\mathit{n}}$gn Conv can serve as a plug-and-play module to improve various vision Transformers and convolution-based models. Based on the proposed operation, we construct a new family of generic vision backbones for various visual modalities and tasks, including HorNet and HorFPN for image recognition, Hor3D for point cloud analysis, and HorCLIP for vision-language modeling. For image recognition, we propose HorNet as a stronger visual encoder, where we conduct extensive experiments on ImageNet classification, COCO object detection, and ADE20K semantic segmentation. HorNet outperforms Swin Transformers and ConvNeXt by a significant margin with similar overall architecture and training configurations. HorNet also shows favorable scalability to more training data and larger model sizes. Apart from image encoders, we also show ${\mathit{g}}^{\mathit{n}}$gnConv can be applied to task-specific decoders and consistently improve dense prediction performance with less computation. For point cloud analysis, we design Hor3D, demonstrating the efficacy of high-order interactions for unstructured point cloud data through experiments on challenging 3D semantic segmentation tasks in S3DIS and ScanNet V2. In vision-language modeling, our proposed HorCLIP surpasses mainstream Vision Transformer and ConvNeXt architectures with shorter training schedules on ImageNet zero-shot classification and shows remarkably higher performance on vision-language dense representation tasks on COCO Panoptic datasets. Our results demonstrate that ${\mathit{g}}^{\mathit{n}}$gnConv with high-order spatial interactions can be a new basic operation for visual modeling that effectively combines the merits of both vision Transformers and CNNs.

Large Language Models (LLMs) demonstrate enhanced capabilities and reliability by reasoning more, evolving from Chain-of-Thought prompting to product-level solutions like OpenAI o1. Despite various efforts to improve LLM reasoning, high-quality long-chain reasoning data and optimized training pipelines still remain inadequately explored in vision-language tasks. In this paper, we present InsightV, an early effort to 1) scalably produce long and robust reasoning data for complex multi-modal tasks, and 2) an effective training pipeline to enhance the reasoning capabilities of multi-modal large language models (MLLMs). Specifically, to create long and structured reasoning data without human labor, we design a two-step pipeline with a progressive strategy to generate sufficiently long and diverse reasoning paths and a multi-granularity assessment method to ensure data quality. We observe that directly supervising MLLMs with such long and complex reasoning data will not yield ideal reasoning ability. To tackle this problem, we design a multi-agent system consisting of a reasoning agent dedicated to performing long-chain reasoning and a summary agent trained to judge and summarize reasoning results. We further incorporate an iterative DPO algorithm to enhance the reasoning agent’s generation stability and quality. Based on the popular LLaVA-NeXT model and our stronger base MLLM, we demonstrate significant performance gains across challenging multi-modal benchmarks requiring visual reasoning. Benefiting from our multi-agent system, Insight-V can also easily maintain or improve performance on perception-focused multi-modal tasks.

Multimodal language models (MLLMs) are increasingly being applied in real-world environments, necessitating their ability to interpret 3D spaces and comprehend temporal dynamics. Current methods often rely on specialized architectural designs or task-specific fine-tuning to achieve this. We introduce Coarse Correspondences, a simple lightweight method that enhances MLLMs’ spatial-temporal reasoning with 2D images as input, without modifying the architecture or requiring task-specific fine-tuning. Our method uses a lightweight tracking model to identify primary object correspondences between frames in a video or across different image viewpoints, and then conveys this information to MLLMs through visual prompting. We demonstrate that this simple training-free approach brings substantial gains to GPT4-V/O consistently on four benchmarks that require spatial-temporal reasoning, including +20.5% improvement on ScanQA, +9.7% on OpenEQA’s episodic memory subset, +6.0% on the long-form video benchmark EgoSchema, and +11% on the R2R navigation benchmark. Additionally, we show that Coarse Correspondences can also enhance open-source MLLMs’ spatial reasoning (by +6.9% on ScanQA) when applied in both training and inference and that the improvement can generalize to unseen datasets such as SQA3D (+3.1%). Taken together, we show that Coarse Correspondences effectively and efficiently boosts models’ performance on downstream tasks requiring spatial-temporal reasoning.

Multimodal Large Language Models (MLLMs) are commonly derived by extending pre-trained Large Language Models (LLMs) with visual capabilities. In this work, we investigate how MLLMs process visual inputs by analyzing their attention mechanisms. We reveal a surprising sparsity phenomenon: only a small subset (approximately less than 5%) of attention heads in LLMs actively contribute to visual understanding, termed visual heads. To identify these heads efficiently, we design a training-free framework that quantifies head-level visual relevance through targeted response analysis. Building on this discovery, we introduce SparseMM, a KV-Cache optimization strategy that allocates asymmetric computation budgets to heads in LLMs based on their visual scores, leveraging the sparity of visual heads for accelerating the inference of MLLMs. Compared with prior KV-Cache acceleration methods that ignore the particularity of visual, SparseMM prioritizes stress and retaining visual semantics during decoding. Extensive evaluations across mainstream multimodal benchmarks demonstrate that SparseMM achieves superior accuracy-efficiency trade-offs. Notably, SparseMM delivers 1.38x real-time acceleration and 52% memory reduction during generation while maintaining performance parity on efficiency test. Our project is open sourced at https://github.com/CR400AF-A/SparseMM.

Visual data comes in various forms, ranging from small icons of just a few pixels to long videos spanning hours. Existing multi-modal LLMs usually standardize these diverse visual inputs to fixed-resolution images or patches for visual encoders and yield similar numbers of tokens for LLMs. This approach is non-optimal for multimodal understanding and inefficient for processing inputs with long and short visual contents. To solve the problem, we propose Oryx, a unified multimodal architecture for the spatial-temporal understanding of images, videos, and multi-view 3D scenes. Oryx offers an on-demand solution to seamlessly and efficiently process visual inputs with arbitrary spatial sizes and temporal lengths through two core innovations: 1) a pre-trained OryxViT model that can encode images at any resolution into LLM-friendly visual representations; 2) a dynamic compressor module that supports 1x to 16x compression on visual tokens by request. These designs enable Oryx to accommodate extremely long visual contexts, such as videos, with lower resolution and high compression while maintaining high recognition precision for tasks like document understanding with native resolution and no compression. Beyond the architectural improvements, enhanced data curation and specialized training on long-context retrieval and spatial-aware data help Oryx achieve strong capabilities in image, video, and 3D multimodal understanding simultaneously.

Reasoning stands as a cornerstone of intelligence, enabling the synthesis of existing knowledge to solve complex problems. Despite remarkable progress, existing reasoning benchmarks often fail to rigorously evaluate the nuanced reasoning capabilities required for complex, real-world problemsolving, particularly in multi-disciplinary and multimodal contexts. In this paper, we introduce a graduate-level, multi-disciplinary, EnglishChinese benchmark, dubbed as Reasoning Bench (RBench), for assessing the reasoning capability of both language and multimodal models. RBench spans 1,094 questions across 108 subjects for language model evaluation and 665 questions across 83 subjects for multimodal model testing. These questions are meticulously curated to ensure rigorous difficulty calibration, subject balance, and cross-linguistic alignment, enabling the assessment to be an Olympiad-level multidisciplinary benchmark. We evaluate many models such as o1, GPT-4o, DeepSeek-R1, etc. Experimental results indicate that advanced models perform poorly on complex reasoning, especially multimodal reasoning. Even the top-performing model OpenAI o1 achieves only 53.2% accuracy on our multimodal evaluation. Data and code are made publicly available athttps://evalmodels.github.io/rbench/

While large language models (LLMs) demonstrate strong reasoning capabilities utilizing reinforcement learning (RL) with verifiable reward, whether large vision-language models (VLMs) can directly inherit such capabilities through similar post-training strategies remains underexplored. In this work, we conduct a systematic compositional probing study to evaluate whether current VLMs trained with RL or other post-training strategies can compose capabilities across modalities or tasks under out-of-distribution conditions. We design a suite of diagnostic tasks that train models on unimodal tasks or isolated reasoning skills, and evaluate them on multimodal, compositional variants requiring skill integration. Through comparisons between supervised fine-tuning (SFT) and RL-trained models, we identify three key findings: (1) RL-trained models consistently outperform SFT on compositional generalization, demonstrating better integration of learned skills; (2) although VLMs achieve strong performance on individual tasks, they struggle to generalize compositionally under cross-modal and cross-task scenarios, revealing a significant gap in current training strategies; (3) enforcing models to explicitly describe visual content before reasoning (e.g., caption-before-thinking), along with rewarding progressive vision-to-text grounding, yields notable gains. It highlights two essential ingredients for improving compositionality in VLMs: visual-to-text alignment and accurate visual grounding. Our findings shed light on the current limitations of RL-based reasoning VLM training and provide actionable insights toward building models that reason compositionally across modalities and tasks.

Unintentional action localization (UAL) is a challenging task that requires reasoning about action intention clues to detect the temporal locations of unintentional action occurrences in real-world videos. Previous efforts usually treated this task as a dense binary classification problem and did not fully explore the relationships between intention clues and unintentional actions, resulting in unsatisfactory performance on open-set scenarios during inference. In this paper, we propose a Transferable Unintentional Action Localization framework by introducing language-guided intention translation, which explicitly formulates unintentional action localization as an open-set localization problem. Our framework constructs a transferable reasoning model guided by natural languages to translate the action intention of the entire video, which generates natural and powerful supervision signals for reconstructing complete action intention clues to address the problem of unintentional action localization. Based on the fact that a video with failure action is composed of intentional and unintentional parts connected by a transient action transition. Our transferable reasoning model employs a transformer architecture to transfer knowledge between intentional and unintentional parts for learning complementary semantic representations of these two parts, completing the action intention clue in an implicit supervision manner. We also present a dense voting scheme for detecting the action transition from intentional to unintentional using discriminative representations incorporating action intention clues. Extensive experiments demonstrate that our framework outperforms representative unintentional action localization methods in a wide range of open-set scenarios. In addition, we create a new unintentional sports video dataset, FS-Falls, and extend our framework from in-the-wild scenarios to competitive sports to demonstrate better generalization ability. We hope this work will provide a new perspective on creating powerful representations with complete action intention priors, which will help us better understand human action and capture underlying intention clues in real-world videos.

Local sampling plays a key role in modeling 3D point clouds. Due to the disordered and unstructured nature of point cloud data, conventional 3D deep models such as PointNet++ and its variants usually employ random or fixed rules to sample local neighborhoods, leading to considerable redundancy in the feature aggregation process. In this paper, we propose a self-supervised method for learning to adaptively select effective neighbors. Firstly, we observe that only a part of sampled points contributes to the aggregated features after the max-pooling operation in existing point cloud models. Then, based on this observation, we propose a simple and task-oriented metric to evaluate the sampling efficiency by measuring the effective neighbors in the feature aggregation process. The metric is also used to supervise a lightweight neighborhood scoring module (NSM), which is designed to efficiently select effective neighboring points from a wider range of neighbors to reduce the computational cost and keep the performance superior. To further improve the performance, we introduce Neighborhood Attention in the feature aggregation process according to the importance score of neighborhood points predicted by NSM. Experimental results show that our method is simple and efficient, and can be applied to most tasks and models to reduce the computational cost and keep the performance superiority. Our code is available athttps://github.com/sunshuofeng/PointMax_Code

Recently, 3D content creation from text prompts has demonstrated remarkable progress by utilizing 2D and 3D diffusion models. While 3D diffusion models ensure great multi-view consistency, their ability to generate highquality and diverse 3D assets is hindered by the limited 3D data. In contrast, 2D diffusion models find a distillation approach that achieves excellent generalization and rich details without any 3D data. However, 2D lifting methods suffer from inherent view-agnostic ambiguity thereby leading to serious multi-face Janus issues, where text prompts fail to provide sufficient guidance to learn coherent 3D results. Instead of retraining a costly viewpoint-aware model, we study how to fully exploit easily accessible coarse 3D knowledge to enhance the prompts and guide 2D lifting optimization for refinement. In this paper, we propose Sherpa3D, a new text-to-3D framework that achieves high-fidelity, generalizability, and geometric consistency simultaneously. Specifically, we design a pair of guiding strategies derived from the coarse 3D prior generated by the 3D diffusion model: a structural guidance for geometric fidelity and a semantic guidance for 3D coherence. Employing the two types of guidance, the 2D diffusion model enriches the 3D content with diversified and high - quality results. Extensive experiments show the superiority of our Sherpa3D over the state-of-the-art text-to-3D methods in terms of quality and 3D consistency. Project page: https://liuff19.github.io/Sherpa3D/.

The human ability to easily solve multimodal tasks in context (i.e., with only a few demonstrations or simple instructions), is what current multimodal systems have largely struggled to imitate. In this work, we demonstrate that the task-agnostic in-context learning capabilities of large multimodal models can be significantly enhanced by effective scaling-up. We introduce Emu2, a generative multimodal model with 37 billion parameters, trained on large-scale multimodal sequences with a unified autoregressive objective. Emu2 exhibits strong multimodal in-context learning abilities, even emerging to solve tasks that require on-the-fly reasoning, such as visual prompting and object-grounded generation. The model sets a new record on multiple multimodal understanding tasks in few-shot settings. When instruction-tuned to follow specific instructions, Emu2 further achieves new state-of-the-art on challenging tasks such as question answering benchmarks for large multimodal models and open-ended subject-driven generation. These achievements demonstrate that Emu2 can serve as a base model and general-purpose interface for a wide range of multimodal tasks. Code and models are publicly available to facilitate future research.

Numerous prior studies predominantly emphasize constructing relation vectors for individual neighborhood points and generating dynamic kernels for each vector and embedding these into high-dimensional spaces to capture implicit local structures. However, we contend that such implicit high-dimensional structure modeling approch inadequately represents the local geometric structure of point clouds due to the absence of explicit structural information. Hence, we introduce X-3D, an explicit 3D structure modeling approach. X-3D functions by capturing the explicit local structural information within the input 3D space and employing it to produce dynamic kernels with shared weights for all neighborhood points within the current local region. This modeling approach introduces effective geometric prior and significantly diminishes the disparity between the local structure of the embedding space and the original input point cloud, thereby improving the extraction of local features. Experiments show that our method can be used on a variety of methods and achieves state-of-the-art performance on segmentation, classification, de-tection tasks with lower extra computational cost, such as 90.7% on ScanObjectNN for classification, 79.2% on S3DIS 6 fold and 74.3% on S3DIS Area 5 for segmentation, 76.3% on ScanNetV2 for segmentation and 64.5% mAP25, 46.9% mAP50on SUN RGB-D and 69.0% mAP25, 51.1% mAP50on ScanNetV2. Our code is available at https://github.com/sunshuofeng/X-3D.

Nowadays, pre-training big models on large-scale datasets has achieved great success and dominated many downstream tasks in natural language processing and 2D vision, while pre-training in 3D vision is still under development. In this paper, we provide a new perspective of transferring the pre-trained knowledge from 2D domain to 3D domain with Point-to-Pixel Prompting in data space and Pixel-to-Point distillation in feature space, exploiting shared knowledge in images and point clouds that display the same visual world. Following the principle of prompting engineering, Point-to-Pixel Prompting transforms point clouds into colorful images with geometry-preserved projection and geometry-aware coloring. Then the pre-trained image models can be directly implemented for point cloud tasks without structural changes or weight modifications. With projection correspondence in feature space, Pixel-to-Point distillation further regards pre-trained image models as the teacher model and distills pre-trained 2D knowledge to student point cloud models, remarkably enhancing inference efficiency and model capacity for point cloud analysis. We conduct extensive experiments in both object classification and scene segmentation under various settings to demonstrate the superiority of our method. In object classification, we reveal the important scale-up trend of Point-to-Pixel Prompting and attain 90.3% accuracy on ScanObjectNN dataset, surpassing previous literature by a large margin. In scene-level semantic segmentation, our method outperforms traditional 3D analysis approaches and shows competitive capacity in dense prediction tasks.

In this paper, we present a new framework named DIML to achieve more interpretable deep metric learning. Unlike traditional deep metric learning method that simply produces a global similarity given two images, DIML computes the overall similarity through the weighted sum of multiple local part-wise similarities, making it easier for human to understand the mechanism of how the model distinguish two images. Specifically, we propose a structural matching strategy that explicitly aligns the spatial embeddings by computing an optimal matching flow between feature maps of the two images. We also devise a multi-scale matching strategy, which considers both global and local similarities and can significantly reduce the computational costs in the application of image retrieval. To handle the view variance in some complicated scenarios, we propose to use cross-correlation as the marginal distribution of the optimal transport to leverage semantic information to locate the important region in the images. Our framework is model-agnostic, which can be applied to off-the-shelf backbone networks and metric learning methods. To extend our DIML to more advanced architectures like vision Transformers (ViTs), we further propose truncated attention rollout and partial similarity to overcome the lack of locality in ViTs. We evaluate our method on three major benchmarks of deep metric learning including CUB200-2011, Cars196, and Stanford Online Products, and achieve substantial improvements over popular metric learning methods with better interpretability.

Person-agnostic face swapping has gained significant attention in recent years, as it offers the potential to enhance various real-world applications by combining high fidelity and identity consistency. However, conventional face swapping methods often rely on intricate adjustments of different loss functions, leading to instability during both the training and inference stages. In this work, we propose a simple yet effective framework namedStableSwapwith a reversible autoencoder to modify the face in a shared latent space. Our approach capitalizes on the information-rich image latent codes to tackle the challenges of complex editing tasks, utilizing the abundant details present in both the source and target faces. To ensure an expressive and robust latent space, we employ a latent alignment approach with perceptual and adversarial losses to optimize the autoencoder. Additionally, we devise a multi-stage identity injection module that samples multiple features with different facial priors and incorporates them to guide the latent image manipulation. By leveraging attention-based blocks, we fuse these futures and update the latent code in a mask-conditioned manner. Both quantitative and qualitative results on the mainstream benchmarks demonstrate that our StableSwap generates competitive identity-consistent swapped faces compared with state-of-the-art methods. Our method outperforms previous approaches in terms of ID Retrieval (98.68) and FID (2.49), while also exhibiting enhanced stability during model training. Beyond this, our model achieves region-controllable face swapping with the capability to perform more fine-grained operations in latent space.

With the continuously thriving popularity around the world, fitness activity analytic has become an emerging research topic in computer vision. While a variety of new tasks and algorithms have been proposed recently, there are growing hunger for data resources involved in high-quality data, fine-grained labels, and diverse environments. In this paper, we present FLAG3D, a large-scale 3D fitness activity dataset with language instruction containing 180K sequences of 60 categories. FLAG3D features the following three aspects: 1) accurate and dense 3D human pose captured from advanced MoCap system to handle the complex activity and large movement, 2) detailed and professional language instruction to describe how to perform a specific activity, 3) versatile video resources from a high-tech MoCap system, rendering software, and cost-effective smartphones in natural environments. Extensive experiments and in-depth analysis show that FLAG3D contributes great research value for various challenges, such as cross-domain human action recognition, dynamic human mesh recovery, and language-guided human action generation. Our dataset and source code are publicly available at https://andytang15.github.io/FLAG3D.

In this paper, we propose DiffSwap, a diffusion model based framework for high-fidelity and controllable face swapping. Unlike previous work that relies on carefully designed network architectures and loss functions to fuse the information from the source and target faces, we reformulate the face swapping as a conditional inpainting task, performed by a powerful diffusion model guided by the desired face attributes (e.g., identity and landmarks). An important issue that makes it nontrivial to apply diffusion models to face swapping is that we cannot perform the time-consuming multi-step sampling to obtain the generated image during training. To overcome this, we propose a mid-point estimation method to efficiently recover a reasonable diffusion result of the swapped face with only 2 steps, which enables us to introduce identity constraints to improve the face swapping quality. Our framework enjoys several favorable properties more appealing than prior arts: 1) Controllable. Our method is based on conditional masked diffusion on the latent space, where the mask and the conditions can be fully controlled and customized. 2) High-fidelity. The formulation of conditional inpainting can fully exploit the generative ability of diffusion models and can preserve the background of target images with minimal artifacts. 3) Shape-preserving. The controllability of our method enables us to use 3D-aware landmarks as the condition during generation to preserve the shape of the source face. Extensive experiments on both FF++ and FFHQ demonstrate that our method can achieve state-of-the-art face swapping results both qualitatively and quantitatively.

With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNet-Part segmentation tasks. Code is available at https://github.com/wangzy22/TakeAPhoto.

In recent years, significant progress has been made in video instance segmentation (VIS), with many offline and online methods achieving state-of-the-art performance. While offline methods have the advantage of producing temporally consistent predictions, they are not suitable for real-time scenarios. Conversely, online methods are more practical, but maintaining temporal consistency remains a challenging task. In this paper, we propose a novel online method for video instance segmentation, called TCOVIS, which fully exploits the temporal information in a video clip. The core of our method consists of a global instance assignment strategy and a spatio-temporal enhancement module, which improve the temporal consistency of the features from two aspects. Specifically, we perform global optimal matching between the predictions and ground truth across the whole video clip, and supervise the model with the global optimal objective. We also capture the spatial feature and aggregate it with the semantic feature between frames, thus realizing the spatio-temporal enhancement. We evaluate our method on four widely adopted VIS benchmarks, namely YouTube-VIS 2019/2021/2022 and OVIS, and achieve state-of-the-art performance on all benchmarks without bells-and-whistles. For instance, on YouTube-VIS 2021, TCOVIS achieves 49.5 AP and 61.3 AP with ResNet-50 and Swin-L backbones, respectively. Code is available at https://github.com/jun-long-li/TCOVIS.

Diffusion models (DMs) have become the new trend of generative models and have demonstrated a powerful ability of conditional synthesis. Among those, text-to-image diffusion models pre-trained on large-scale image-text pairs are highly controllable by customizable prompts. Unlike the unconditional generative models that focus on low-level attributes and details, text-to-image diffusion models contain more high-level knowledge thanks to the vision-language pre-training. In this paper, we propose VPD (Visual Perception with pre-trained Diffusion models), a new framework that exploits the semantic information of a pre-trained text-to-image diffusion model in visual perception tasks. Instead of using the pre-trained denoising autoencoder in a diffusion-based pipeline, we simply use it as a backbone and aim to study how to take full advantage of the learned knowledge. Specifically, we prompt the denoising decoder with proper textual inputs and refine the text features with an adapter, leading to a better alignment to the pre-trained stage and making the visual contents interact with the text prompts. We also propose to utilize the cross-attention maps between the visual features and the text features to provide explicit guidance. Compared with other pre-training methods, we show that vision-language pre-trained diffusion models can be faster adapted to downstream visual perception tasks using the proposed VPD. Extensive experiments on semantic segmentation, referring image segmentation, and depth estimation demonstrate the effectiveness of our method. Notably, VPD attains 0.254 RMSE on NYUv2 depth estimation and 73.3% oIoU on RefCOCO-val referring image segmentation, establishing new records on these two benchmarks. Code is available at https://github.com/wl-zhao/VPD.

Diffusion probabilistic models (DPMs) have demonstrated a very promising ability in high-resolution image synthesis. However, sampling from a pre-trained DPM is time-consuming due to the multiple evaluations of the denoising network, making it more and more important to accelerate the sampling of DPMs. Despite recent progress in designing fast samplers, existing methods still cannot generate satisfying images in many applications where fewer steps (e.g., $<$10) are favored. In this paper, we develop a unified corrector (UniC) that can be applied after any existing DPM sampler to increase the order of accuracy without extra model evaluations, and derive a unified predictor (UniP) that supports arbitrary order as a byproduct. Combining UniP and UniC, we propose a unified predictor-corrector framework called UniPC for the fast sampling of DPMs, which has a unified analytical form for any order and can significantly improve the sampling quality over previous methods, especially in extremely few steps. We evaluate our methods through extensive experiments including both unconditional and conditional sampling using pixel-space and latent-space DPMs. Our UniPC can achieve 3.87 FID on CIFAR10 (unconditional) and 7.51 FID on ImageNet 256$\times$256 (conditional) with only 10 function evaluations. Code is available at https://github.com/wl-zhao/UniPC.

This work explores the use of global and local structures of 3D point clouds as a free and powerful supervision signal for representation learning. Local and global patterns of a 3D object are closely related. Although each part of an object is incomplete, the underlying attributes about the object are shared among all parts, which makes reasoning about the whole object from a single part possible. We hypothesize that a powerful representation of a 3D object should model the attributes that are shared between parts and the whole object, and distinguishable from other objects. Based on this hypothesis, we propose a new framework to learn point cloud representations by bidirectional reasoning between the local structures at different abstraction hierarchies and the global shape. Moreover, we extend the unsupervised structural representation learning method to more complex 3D scenes. By introducing structural proxies as the intermediate-level representations between local and global ones, we propose a hierarchical reasoning scheme among local parts, structural proxies, and the overall point cloud to learn powerful 3D representations in an unsupervised manner. Extensive experimental results demonstrate that the unsupervised representations can be very competitive alternatives of supervised representations in discriminative power, and exhibit better performance in generalization ability and robustness. Our method establishes the new state-of-the-art of unsupervised/few-shot 3D object classification and part segmentation. We also show our method can serve as a simple yet effective regime for model pre-training on 3D scene segmentation and detection tasks. We expect our observations to offer a new perspective on learning better representations from data structures instead of human annotations for point cloud understanding.

In this paper, we present a new approach for model acceleration by exploiting spatial sparsity in visual data. We observe that the final prediction in vision Transformers is only based on a subset of the most informative regions, which is sufficient for accurate image recognition. Based on this observation, we propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically based on the input to accelerate vision Transformers. Specifically, we devise a lightweight prediction module to estimate the importance of each token given the current features. The module is added to different layers to prune redundant tokens hierarchically. While the framework is inspired by our observation of the sparse attention in vision Transformers, we find that the idea of adaptive and asymmetric computation can be a general solution for accelerating various architectures. We extend our method to hierarchical models including CNNs and hierarchical vision Transformers as well as more complex dense prediction tasks. To handle structured feature maps, we formulate a generic dynamic spatial sparsification framework with progressive sparsification and asymmetric computation for different spatial locations. By applying lightweight fast paths to less informative features and expressive slow paths to important locations, we can maintain the complete structure of feature maps while significantly reducing the overall computations. Extensive experiments on diverse modern architectures and different visual tasks demonstrate the effectiveness of our proposed framework. By hierarchically pruning 66% of the input tokens, our method greatly reduces 31% ∼ 35% FLOPs and improves the throughput by over 40% while the drop of accuracy is within 0.5% for various vision Transformers. By introducing asymmetric computation, a similar acceleration can be achieved on modern CNNs and Swin Transformers. Moreover, our method achieves promising results on more complex tasks including semantic segmentation and object detection. Our results clearly demonstrate that dynamic spatial sparsification offers a new and more effective dimension for model acceleration. Code is available at https://github.com/raoyongming/DynamicViT.

Recent advances in self-attention and pure multi-layer perceptrons (MLP) models for vision have shown great potential in achieving promising performance with fewer inductive biases. These models are generally based on learning interaction among spatial locations from raw data. The complexity of self-attention and MLP grows quadratically as the image size increases, which makes these models hard to scale up when high-resolution features are required. In this paper, we present the Global Filter Network (GFNet), a conceptually simple yet computationally efficient architecture, that learns long-term spatial dependencies in the frequency domain with log-linear complexity. Our architecture replaces the self-attention layer in vision Transformers with three key operations: a 2D discrete Fourier transform, an element-wise multiplication between frequency-domain features and learnable global filters, and a 2D inverse Fourier transform. Based on this basic design, we develop a series of isotropic models with a Transformer-style simple architecture and CNN-style hierarchical models with better performance. Isotropic GFNet models exhibit favorable accuracy/complexity trade-offs compared to recent vision Transformers and pure MLP models. Hierarchical GFNet models can inherit successful designs in CNNs and be easily scaled up with larger model sizes and more training data, showing strong performance on both image classification (e.g., 85.0% top-1 accuracy on ImageNet-1 k without any extra data or supervision, and 87.4% accuracy with ImageNet-21 k pre-training) and dense prediction tasks (e.g., 54.3 mIoU on ADE20 k val). Our results demonstrate that GFNet can be a very competitive alternative to Transformer-based models and CNNs in terms of efficiency, generalization ability and robustness. Code is available at https://github.com/raoyongming/GFNet.

In this paper, we propose Point-Voxel Correlation Fields to explore relations between two consecutive point clouds and estimate scene flow that represents 3D motions. Most existing works only consider local correlations, which are able to handle small movements but fail when there are large displacements. Therefore, it is essential to introduce all-pair correlation volumes that are free from local neighbor restrictions and cover both short- and long-term dependencies. However, it is challenging to efficiently extract correlation features from all-pairs fields in the 3D space, given the irregular and unordered nature of point clouds. To tackle this problem, we present point-voxel correlation fields, proposing distinct point and voxel branches to inquire about local and long-range correlations from all-pair fields respectively. To exploit point-based correlations, we adopt the K-Nearest Neighbors search that preserves fine-grained information in the local region, which guarantees the scene flow estimation precision. By voxelizing point clouds in a multi-scale manner, we construct pyramid correlation voxels to model long-range correspondences, which are utilized to handle fast-moving objects. Integrating these two types of correlations, we propose Point-Voxel Recurrent All-Pairs Field Transforms (PV-RAFT) architecture that employs an iterative scheme to estimate scene flow from point clouds. To adapt to different flow scope conditions and obtain more fine-grained results, we further propose Deformable PV-RAFT (DPV-RAFT), where the Spatial Deformation deforms the voxelized neighborhood, and the Temporal Deformation controls the iterative update process. We evaluate the proposed method on the FlyingThings3D and KITTI Scene Flow 2015 datasets and experimental results show that we outperform state-of-the-art methods by remarkable margins.

In this paper, we propose a Transformer encoder-decoder architecture, called PoinTr, which reformulates point cloud completion as a set-to-set translation problem and employs a geometry-aware block to model local geometric relationships explicitly. The migration of Transformers enables our model to better learn structural knowledge and preserve detailed information for point cloud completion. Taking a step towards more complicated and diverse situations, we further propose AdaPoinTr by developing an adaptive query generation mechanism and designing a novel denoising task during completing a point cloud. Coupling these two techniques enables us to train the model efficiently and effectively: we reduce training time (by 15x or more) and improve completion performance (over 20%). Additionally, we propose two more challenging benchmarks with more diverse incomplete point clouds that can better reflect real-world scenarios to promote future research. We also show our method can be extended to the scene-level point cloud completion scenario by designing a new geometry-enhanced semantic scene completion framework. Extensive experiments on the existing and newly-proposed datasets demonstrate the effectiveness of our method, which attains 6.53 CD on PCN, 0.81 CD on ShapeNet-55 and 0.392 MMD on real-world KITTI, surpassing other work by a large margin and establishing new state-of-the-arts on various benchmarks. Most notably, AdaPoinTr can achieve such promising performance with higher throughputs and fewer FLOPs compared with the previous best methods in practice.

Human behavior has the nature of indeterminacy, which requires the pedestrian trajectory prediction system to model the multi-modality of future motion states. Unlike existing stochastic trajectory prediction methods which usually use a latent variable to represent multi-modality, we explicitly simulate the process of human motion variation from indeterminate to determinate. In this paper, we present a new framework to formulate the trajectory prediction task as a reverse process of motion indeterminacy diffusion (MID), in which we progressively discard indeterminacy from all the walkable areas until reaching the desired trajectory. This process is learned with a parameterized Markov chain conditioned by the observed trajectories. We can adjust the length of the chain to control the degree of indeterminacy and balance the diversity and determinacy of the predictions. Specifically, we encode the history behavior information and the social interactions as a state embedding and devise a Transformer-based diffusion model to capture the temporal dependencies of trajectories. Extensive experiments on the human trajectory prediction benchmarks including the Stanford Drone and ETH/UCY datasets demonstrate the superiority of our method. Code is available at https://github.com/gutianpei/MID.

Recent progress has shown that large-scale pre-training using contrastive image-text pairs can be a promising alternative for high-quality visual representation learning from natural language supervision. Benefiting from a broader source of supervision, this new paradigm exhibits impressive transferability to downstream classification tasks and datasets. However, the problem of transferring the knowledge learned from image-text pairs to more complex dense prediction tasks has barely been visited. In this work, we present a new framework for dense prediction by implicitly and explicitly leveraging the pre-trained knowledge from CLIP. Specifically, we convert the original image-text matching problem in CLIP to a pixel-text matching problem and use the pixel-text score maps to guide the learning of dense prediction models. By further using the contextual information from the image to prompt the language model, we are able to facilitate our model to better exploit the pretrained knowledge. Our method is model-agnostic, which can be applied to arbitrary dense prediction systems and various pre-trained visual backbones including both CLIP models and ImageNet pre-trained models. Extensive experiments demonstrate the superior performance of our methods on semantic segmentation, object detection, and instance segmentation tasks. Code is available at https://github.com/raoyongming/DenseCLIP.

Conventional point cloud semantic segmentation methods usually employ an encoder-decoder architecture, where mid-level features are locally aggregated to extract geometric information. However, the over-reliance on these class-agnostic local geometric representations may raise confusion between local parts from different categories that are similar in appearance or spatially adjacent. To address this issue, we argue that mid-level features can be further enhanced with semantic information, and propose semantic-affine transformation that transforms features of mid-level points belonging to different categories with class-specific affine parameters. Based on this technique, we propose SemAffiNet for point cloud semantic segmentation, which utilizes the attention mechanism in the Transformer module to implicitly and explicitly capture global structural knowledge within local parts for overall comprehension of each category. We conduct extensive experiments on the ScanNetV2 and NYUv2 datasets, and evaluate semantic-affine transformation on various 3D point cloud and 2D image segmentation baselines, where both qualitative and quantitative results demonstrate the superiority and generalization ability of our proposed approach. Code is available at https://github.com/wangzy22/SemAffiNet.

Most existing action quality assessment methods rely on the deep features of an entire video to predict the score, which is less reliable due to the non-transparent inference process and poor interpretability. We argue that understanding both high-level semantics and internal temporal structures of actions in competitive sports videos is the key to making predictions accurate and interpretable. Towards this goal, we construct a new fine-grained dataset, called FineDiving, developed on diverse diving events with detailed annotations on action procedures. We also propose a procedure-aware approach for action quality assessment, learned by a new Temporal Segmentation Attention module. Specifically, we propose to parse pairwise query and exemplar action instances into consecutive steps with diverse semantic and temporal correspondences. The procedure-aware cross-attention is proposed to learn embeddings between query and exemplar steps to discover their semantic, spatial, and temporal correspondences, and further serve for fine-grained contrastive regression to derive a reliable scoring mechanism. Extensive experiments demonstrate that our approach achieves substantial improvements over the state-of-the-art methods with better interpretability. The dataset and code are available at https://github.com/xujinglin/FineDiving.

In this paper, we propose a weakly-supervised approach for 3D object detection, which makes it possible to train a strong 3D detector with position-level annotations (i.e. annotations of object centers). In order to remedy the information loss from box annotations to centers, our method, namely Back to Reality (BR), makes use of synthetic 3D shapes to convert the weak labels into fully-annotated virtual scenes as stronger supervision, and in turn utilizes the perfect virtual labels to complement and refine the real labels. Specifically, we first assemble 3D shapes into physically reasonable virtual scenes according to the coarse scene layout extracted from position-level annotations. Then we go back to reality by applying a virtual-to-real domain adaptation method, which refine the weak labels and additionally supervise the training of detector with the virtual scenes. Furthermore, we propose a more challenging benckmark for indoor 3D object detection with more diversity in object sizes for better evaluation. With less than 5% of the labeling labor, we achieve comparable detection performance with some popular fully-supervised approaches on the widely used ScanNet dataset. Code is available at: https://github.com/wyf-ACCEPT/BackToReality.

We present Point-BERT, a new paradigm for learning Transformers to generalize the concept of BERT [8] to 3D point cloud. Inspired by BERT, we devise a Masked Point Modeling (MPM) task to pre-train point cloud Transformers. Specifically, we first divide a point cloud into several local point patches, and a point cloud Tokenizer with a discrete Variational AutoEncoder (dVAE) is designed to generate discrete point tokens containing meaningful local information. Then, we randomly mask out some patches of input point clouds and feed them into the backbone Transformers. The pre-training objective is to recover the original point tokens at the masked locations under the supervision of point tokens obtained by the Tokenizer. Extensive experiments demonstrate that the proposed BERT-style pre-training strategy significantly improves the performance of standard point cloud Transformers. Equipped with our pre-training strategy, we show that a pure Transformer architecture attains 93.8% accuracy on ModelNet40 and 83.1% accuracy on the hardest setting of ScanObjectNN, surpassing carefully designed point cloud models with much fewer hand-made designs. We also demonstrate that the representations learned by Point-BERT transfer well to new tasks and domains, where our models largely advance the state-of-the-art of few-shot point cloud classification task. The code and pre-trained models are available at https://github.com/lulutang0608/Point-BERT.

Recent progress in vision Transformers exhibits great success in various tasks driven by the new spatial modeling mechanism based on dot-product self-attention. In this paper, we show that the key ingredients behind the vision Transformers, namely input-adaptive, long-range and high-order spatial interactions, can also be efficiently implemented with a convolution-based framework. We present the Recursive Gated Convolution ($\textit{g}^\textit{n}$Conv) that performs high-order spatial interactions with gated convolutions and recursive designs. The new operation is highly flexible and customizable, which is compatible with various variants of convolution and extends the two-order interactions in self-attention to arbitrary orders without introducing significant extra computation. $\textit{g}^\textit{n}$Conv can serve as a plug-and-play module to improve various vision Transformers and convolution-based models. Based on the operation, we construct a new family of generic vision backbones named HorNet. Extensive experiments on ImageNet classification, COCO object detection and ADE20K semantic segmentation show HorNet outperform Swin Transformers and ConvNeXt by a significant margin with similar overall architecture and training configurations. HorNet also shows favorable scalability to more training data and larger model sizes. Apart from the effectiveness in visual encoders, we also show $\textit{g}^\textit{n}$Conv can be applied to task-specific decoders and consistently improve dense prediction performance with less computation. Our results demonstrate that $\textit{g}^\textit{n}$Conv can be a new basic module for visual modeling that effectively combines the merits of both vision Transformers and CNNs. Code is available at https://github.com/raoyongming/HorNet.

Nowadays, pre-training big models on large-scale datasets has become a crucial topic in deep learning. The pre-trained models with high representation ability and transferability achieve a great success and dominate many downstream tasks in natural language processing and 2D vision. However, it is non-trivial to promote such a pretraining-tuning paradigm to the 3D vision, given the limited training data that are relatively inconvenient to collect. In this paper, we provide a new perspective of leveraging pre-trained 2D knowledge in 3D domain to tackle this problem, tuning pre-trained image models with the novel Point-to-Pixel prompting for point cloud analysis at a minor parameter cost. Following the principle of prompting engineering, we transform point clouds into colorful images with geometry-preserved projection and geometry-aware coloring to adapt to pre-trained image models, whose weights are kept frozen during the end-to-end optimization of point cloud analysis tasks. We conduct extensive experiments to demonstrate that cooperating with our proposed Point-to-Pixel Prompting, better pre-trained image model will lead to consistently better performance in 3D vision. Enjoying prosperous development from image pre-training field, our method attains 89.3% accuracy on the hardest setting of ScanObjectNN, surpassing conventional point cloud models with much fewer trainable parameters. Our framework also exhibits very competitive performance on ModelNet classification and ShapeNet Part Segmentation. Code is available at https://github.com/wangzy22/P2P.

Structures matter in single image super-resolution (SISR). Benefiting from generative adversarial networks (GANs), recent studies have promoted the development of SISR by recovering photo-realistic images. However, there are still undesired structural distortions in the recovered images. In this paper, we propose a structure-preserving super-resolution (SPSR) method to alleviate the above issue while maintaining the merits of GAN-based methods to generate perceptual-pleasant details. First, we propose SPSR with gradient guidance (SPSR-G) by exploiting gradient maps of images to guide the recovery in two aspects. On the one hand, we restore high-resolution gradient maps by a gradient branch to provide additional structure priors for the SR process. On the other hand, we propose a gradient loss to impose a second-order restriction on the super-resolved images, which helps generative networks concentrate more on geometric structures. Second, since the gradient maps are handcrafted and may only be able to capture limited aspects of structural information, we further extend SPSR-G by introducing a learnable neural structure extractor (NSE) to unearth richer local structures and provide stronger supervision for SR. We propose two self-supervised structure learning methods, contrastive prediction and solving jigsaw puzzles, to train the NSEs. Our methods are model-agnostic, which can be potentially used for off-the-shelf SR networks. Experimental results on five benchmark datasets show that the proposed methods outperform state-of-the-art perceptual-driven SR methods under LPIPS, PSNR, and SSIM metrics. Visual results demonstrate the superiority of our methods in restoring structures while generating natural SR images. Code is available at https://github.com/Maclory/SPSR.

Video saliency prediction (VSP) aims to imitate eye fixations of humans. However, the potential of this task has not been fully exploited since existing VSP methods only focus on modeling visual saliency of the input previous frames. In this paper, we present the first attempt to extend this task to video saliency forecasting (VSF) by forecasting attention regions of consecutive future frames. To tackle this problem, we propose a video saliency forecasting transformer (VSFT) network built on a new encoder-decoder architecture. Different from existing VSP methods, our VSFT is the first pure-transformer based architecture in the VSP field and is freed from the dependency of the pretrained S3D model. In VSFT, the attention mechanism is exploited to capture spatial-temporal dependencies between the input past frames and the target future frame. We propose cross-attention guidance blocks (CAGB) to aggregate multi-level representation features to provide sufficient guidance for forecasting. We conduct comprehensive experiments on two benchmark datasets, DHF1K and Hollywoods-2. We investigate the saliency forecasting and predicting abilities of existing VSP methods by modifying the supervision signals. Experimental results demonstrate that our method achieves superior performance on both VSF and VSP tasks.

In this paper, we investigate the problem of abductive visual reasoning (AVR), which requires vision systems to infer the most plausible explanation for visual observations. Unlike previous work which performs visual reasoning on static images or synthesized scenes, we exploit long-term reasoning from instructional videos that contain a wealth of detailed information about the physical world. We conceptualize two tasks for this emerging and challenging topic. The primary task is AVR, which is based on the initial configuration and desired goal from an instructional video, and the model is expected to figure out what is the most plausible sequence of steps to achieve the goal. In order to avoid trivial solutions based on appearance information rather than reasoning, the second task called AVR++ is constructed, which requires the model to answer why the unselected options are less plausible. We introduce a new dataset called VideoABC, which consists of 46,354 unique steps derived from 11,827 instructional videos, formulated as 13,526 abductive reasoning questions with an average reasoning duration of 51 seconds. Through an adversarial hard hypothesis mining algorithm, non-trivial and high-quality problems are generated efficiently and effectively. To achieve human-level reasoning, we propose a Hierarchical Dual Reasoning Network (HDRNet) to capture the long-term dependencies among steps and observations. We establish a benchmark for abductive visual reasoning, and our method set state-of-the-arts on AVR ( ∼ 74 %) and AVR++ ( ∼ 45 %), and humans can easily achieve over 90% accuracy on these two tasks. The large performance gap reveals the limitation of current video understanding models on temporal reasoning and leaves substantial room for future research on this challenging problem. Our dataset and code are available at https://github.com/wl-zhao/VideoABC.

Most widely-used convolutional neural networks (CNNs) end up with a global average pooling layer and a fully-connected layer. In this pipeline, a certain class is represented by one template vector preserved in the feature banks of fully-connected layer. Yet, a class may have multiple properties useful for recognition while the above formulation only captures one of them. Therefore, it is desired to represent a class by multiple proxies. However, directly adding multiple linear layers turns out to be a trivial solution as no improvement can be observed. To tackle this problem, we adopt optimal transport theory to calculate a non-uniform matching flow between the elements in the feature map of a sample and the proxies of a class in a closed way. By doing so, the models are enabled to achieve partial matching as both the feature maps and the proxy set can now focus on a subset of elements from the counterpart. Such formulation also enables us to embed the samples into the Wasserstein metric space, which has many advantages over the original Euclidean space. This formulation can be achieved by a lightweight iterative algorithm, which can be easily embedded into the automatic differentiation framework. Empirical studies are performed on two widely-used classification datasets, CIFAR, and ILSVRC2012, and the substantial improvements on these two benchmarks demonstrate the effectiveness of our method.

In this paper, we propose a Point-Voxel Recurrent All-Pairs Field Transforms (PV-RAFT) method to estimate scene flow from point clouds. Since point clouds are irregular and unordered, it is challenging to efficiently extract features from all-pairs fields in the 3D space, where all-pairs correlations play important roles in scene flow estimation. To tackle this problem, we present point-voxel correlation fields, which capture both local and long-range dependencies of point pairs. To capture point-based correlations, we adopt the K-Nearest Neighbors search that preserves fine-grained information in the local region. By voxelizing point clouds in a multi-scale manner, we construct pyramid correlation voxels to model long-range correspondences. Integrating these two types of correlations, our PV-RAFT makes use of all-pairs relations to handle both small and large displacements. We evaluate the proposed method on the FlyingThings3D and KITTI Scene Flow 2015 datasets. Experimental results show that PV-RAFT outperforms state-of-the-art methods by remarkable margins.

Attention mechanism has demonstrated great potential in fine-grained visual recognition tasks. In this paper, we present a counterfactual attention learning method to learn more effective attention based on causal inference. Unlike most existing methods that learn visual attention based on conventional likelihood, we propose to learn the attention with counterfactual causality, which provides a tool to measure the attention quality and a powerful supervisory signal to guide the learning process. Specifically, we analyze the effect of the learned visual attention on network prediction through counterfactual intervention and maximize the effect to encourage the network to learn more useful attention for fine-grained image recognition. Empirically, we evaluate our method on a wide range of fine-grained recognition tasks where attention plays a crucial role, including fine-grained image categorization, person re-identification, and vehicle re-identification. The consistent improvement on all benchmarks demonstrates the effectiveness of our method.1

3D point cloud understanding has made great progress in recent years. However, one major bottleneck is the scarcity of annotated real datasets, especially compared to 2D object detection tasks, since a large amount of labor is involved in annotating the real scans of a scene. A promising solution to this problem is to make better use of the synthetic dataset, which consists of CAD object models, to boost the learning on real datasets. This can be achieved by the pre-training and fine-tuning procedure. However, recent work on 3D pre-training exhibits failure when transfer features learned on synthetic objects to other real-world applications. In this work, we put forward a new method called RandomRooms to accomplish this objective. In particular, we propose to generate random layouts of a scene by making use of the objects in the synthetic CAD dataset and learn the 3D scene representation by applying object-level contrastive learning on two random scenes generated from the same set of synthetic objects. The model pre-trained in this way can serve as a better initialization when later fine-tuning on the 3D object detection task. Empirically, we show consistent improvement in downstream 3D detection tasks on several base models, especially when less training data are used, which strongly demonstrates the effectiveness and generalization of our method. Benefiting from the rich semantic knowledge and diverse objects from synthetic data, our method establishes the new state-of-the-art on widely-used 3D detection benchmarks ScanNetV2 and SUN RGB-D. We expect our attempt to provide a new perspective for bridging object and scene-level 3D understanding.

In this work, we present a new multi-view depth estimation method that utilizes both conventional SfM reconstruction and learning-based priors over the recently proposed neural radiance fields (NeRF). Unlike existing neural network based optimization method that relies on estimated correspondences, our method directly optimizes over implicit volumes, eliminating the challenging step of matching pixels in indoor scenes. The key to our approach is to utilize the learning-based priors to guide the optimization process of NeRF. Our system firstly adapts a monocular depth network over the target scene by finetuning on its sparse SfM reconstruction. Then, we show that the shape-radiance ambiguity of NeRF still exists in indoor environments and propose to address the issue by employing the adapted depth priors to monitor the sampling process of volume rendering. Finally, a per-pixel confidence map acquired by error computation on the rendered image can be used to fur ther improve the depth quality. Experiments show that our proposed framework significantly outperforms state-of-the-art methods on indoor scenes, with surprising findings presented on the effectiveness of correspondence-based opti-mization and NeRF-based optimization over the adapted depth priors. In addition, we show that the guided opti-mization scheme does not sacrifice the original synthesis capability of neural radiance fields, improving the rendering quality on both seen and novel views. Code is available at https://github.com/weiyithu/NerfingMVS.

Point clouds captured in real-world applications are of-ten incomplete due to the limited sensor resolution, single viewpoint, and occlusion. Therefore, recovering the complete point clouds from partial ones becomes an indispensable task in many practical applications. In this paper, we present a new method that reformulates point cloud completion as a set-to-set translation problem and design a new model, called PoinTr that adopts a transformer encoder-decoder architecture for point cloud completion. By rep-resenting the point cloud as a set of unordered groups of points with position embeddings, we convert the point cloud to a sequence of point proxies and employ the transformers for point cloud generation. To facilitate transformers to better leverage the inductive bias about 3D geometric structures of point clouds, we further devise a geometry-aware block that models the local geometric relationships explicitly. The migration of transformers enables our model to better learn structural knowledge and preserve detailed information for point cloud completion. Furthermore, we propose two more challenging benchmarks with more diverse incomplete point clouds that can better reflect the real-world scenarios to promote future research. Experimental results show that our method outperforms state-of-the-art methods by a large margin on both the new bench-marks and the existing ones. Code is available at https://github.com/yuxumin/PoinTr.

Assessing action quality is challenging due to the subtle differences between videos and large variations in scores. Most existing approaches tackle this problem by regressing a quality score from a single video, suffering a lot from the large inter-video score variations. In this paper, we show that the relations among videos can provide important clues for more accurate action quality assessment during both training and inference. Specifically, we reformulate the problem of action quality assessment as regressing the relative scores with reference to another video that has shared attributes (e.g., category and difficulty), instead of learning unreferenced scores. Following this formulation, we propose a new Contrastive Regression (CoRe) framework to learn the relative scores by pair-wise comparison, which highlights the differences between videos and guides the models to learn the key hints for assessment. In order to further exploit the relative information between two videos, we devise a group-aware regression tree to convert the conventional score regression into two easier sub-problems: coarse-to-fine classification and regression in small intervals. To demonstrate the effectiveness of CoRe, we conduct extensive experiments on three mainstream AQA datasets including AQA-7, MTL-AQA and JIGSAWS. Our approach outperforms previous methods by a large margin and establishes new state-of-the-art on all three benchmarks.

How do the neural networks distinguish two images? It is of critical importance to understand the matching mechanism of deep models for developing reliable intelligent systems for many risky visual applications such as surveillance and access control. However, most existing deep metric learning methods match the images by comparing feature vectors, which ignores the spatial structure of images and thus lacks interpretability. In this paper, we present a deep interpretable metric learning (DIML) method for more transparent embedding learning. Unlike conventional metric learning methods based on feature vector comparison, we propose a structural matching strategy that explicitly aligns the spatial embeddings by computing an optimal matching flow between feature maps of the two images. Our method enables deep models to learn metrics in a more human-friendly way, where the similarity of two images can be decomposed to several part-wise similarities and their contributions to the overall similarity. Our method is model-agnostic, which can be applied to off-the-shelf backbone networks and metric learning methods. We evaluate our method on three major benchmarks of deep metric learning including CUB200-2011, Cars196, and Stanford Online Products, and achieve substantial improvements over popular metric learning methods with better interpretability. Code is available at https://github.com/wl-zhao/DIML.

Attention is sparse in vision transformers. We observe the final prediction in vision transformers is only based on a subset of most informative tokens, which is sufficient for accurate image recognition. Based on this observation, we propose a dynamic token sparsification framework to prune redundant tokens progressively and dynamically based on the input. Specifically, we devise a lightweight prediction module to estimate the importance score of each token given the current features. The module is added to different layers to prune redundant tokens hierarchically. To optimize the prediction module in an end-to-end manner, we propose an attention masking strategy to differentiably prune a token by blocking its interactions with other tokens. Benefiting from the nature of self-attention, the unstructured sparse tokens are still hardware friendly, which makes our framework easy to achieve actual speed-up. By hierarchically pruning 66% of the input tokens, our method greatly reduces 31% $\sim$ 37% FLOPs and improves the throughput by over 40% while the drop of accuracy is within 0.5% for various vision transformers. Equipped with the dynamic token sparsification framework, DynamicViT models can achieve very competitive complexity/accuracy trade-offs compared to state-of-the-art CNNs and vision transformers on ImageNet. Code is available at https://github.com/raoyongming/DynamicViT

Recent advances in self-attention and pure multi-layer perceptrons (MLP) models for vision have shown great potential in achieving promising performance with fewer inductive biases. These models are generally based on learning interaction among spatial locations from raw data. The complexity of self-attention and MLP grows quadratically as the image size increases, which makes these models hard to scale up when high-resolution features are required. In this paper, we present the Global Filter Network (GFNet), a conceptually simple yet computationally efficient architecture, that learns long-term spatial dependencies in the frequency domain with log-linear complexity. Our architecture replaces the self-attention layer in vision transformers with three key operations: a 2D discrete Fourier transform, an element-wise multiplication between frequency-domain features and learnable global filters, and a 2D inverse Fourier transform. We exhibit favorable accuracy/complexity trade-offs of our models on both ImageNet and downstream tasks. Our results demonstrate that GFNet can be a very competitive alternative to transformer-style models and CNNs in efficiency, generalization ability and robustness. Code is available at https://github.com/raoyongming/GFNet

Recent works based on deep learning and facial priors have succeeded in super-resolving severely degraded facial images. However, the prior knowledge is not fully exploited in existing methods, since facial priors such as landmark and component maps are always estimated by low-resolution or coarsely super-resolved images, which may be inaccurate and thus affect the recovery performance. In this paper, we propose a deep face super-resolution (FSR) method with iterative collaboration between two recurrent networks which focus on facial image recovery and landmark estimation respectively. In each recurrent step, the recovery branch utilizes the prior knowledge of landmarks to yield higher-quality images which facilitate more accurate landmark estimation in turn. Therefore, the iterative information interaction between two processes boosts the performance of each other progressively. Moreover, a new attentive fusion module is designed to strengthen the guidance of landmark maps, where facial components are generated individually and aggregated attentively for better restoration. Quantitative and qualitative experimental results show the proposed method significantly outperforms state-of-the-art FSR methods in recovering high-quality face images.

Structures matter in single image super resolution (SISR). Recent studies benefiting from generative adversarial network (GAN) have promoted the development of SISR by recovering photo-realistic images. However, there are always undesired structural distortions in the recovered images. In this paper, we propose a structure-preserving super resolution method to alleviate the above issue while maintaining the merits of GAN-based methods to generate perceptual-pleasant details. Specifically, we exploit gradient maps of images to guide the recovery in two aspects. On the one hand, we restore high-resolution gradient maps by a gradient branch to provide additional structure priors for the SR process. On the other hand, we propose a gradient loss which imposes a second-order restriction on the super-resolved images. Along with the previous image-space loss functions, the gradient-space objectives help generative networks concentrate more on geometric structures. Moreover, our method is model-agnostic, which can be potentially used for off-the-shelf SR networks. Experimental results show that we achieve the best PI and LPIPS performance and meanwhile comparable PSNR and SSIM compared with state-of-the-art perceptual-driven SR methods. Visual results demonstrate our superiority in restoring structures while generating natural SR images.

Local and global patterns of an object are closely related. Although each part of an object is incomplete, the underlying attributes about the object are shared among all parts, which makes reasoning the whole object from a single part possible. We hypothesize that a powerful representation of a 3D object should model the attributes that are shared between parts and the whole object, and distinguishable from other objects. Based on this hypothesis, we propose to learn point cloud representation by bidirectional reasoning between the local structures at different abstraction hierarchies and the global shape without human supervision. Experimental results on various benchmark datasets demonstrate the unsupervisedly learned representation is even better than supervised representation in discriminative power, generalization ability, and robustness. We show that unsupervisedly trained point cloud models can outperform their supervised counterparts on downstream classification tasks. Most notably, by simply increasing the channel width of an SSG PointNet++, our unsupervised model surpasses the state-of-the-art supervised methods on both synthetic and real-world 3D object classification datasets. We expect our observations to offer a new perspective on learning better representation from data structures instead of human annotations for point cloud understanding.

We present a generic, flexible and 3D rotation invariant framework based on spherical symmetry for point cloud recognition. By introducing regular icosahedral lattice and its fractals to approximate and discretize sphere, convolution can be easily implemented to process 3D points. Based on the fractal structure, a hierarchical feature learning framework together with an adaptive sphere projection module is proposed to learn deep feature in an end-to-end manner. Our framework not only inherits the strong representation power and generalization capability from convolutional neural networks for image recognition, but also extends CNN to learn robust feature resistant to rotations and perturbations. The proposed model is effective yet robust. Comprehensive experimental study demonstrates that our approach can achieve competitive performance compared to state-of-the-art techniques on both 3D object classification and part segmentation tasks, meanwhile, outperform other rotation invariant models on rotated 3D object classification and retrieval tasks by a large margin.

There are substantial instruction videos on the Internet, which enables us to acquire knowledge for completing various tasks. However, most existing datasets for instruction video analysis have the limitations in diversity and scale, which makes them far from many real-world applications where more diverse activities occur. Moreover, it still remains a great challenge to organize and harness such data. To address these problems, we introduce a large-scale dataset called “COIN" for COmprehensive INstruction video analysis. Organized with a hierarchical structure, the COIN dataset contains 11,827 videos of 180 tasks in 12 domains (e.g., vehicles, gadgets, etc.) related to our daily life. With a new developed toolbox, all the videos are annotated effectively with a series of step descriptions and the corresponding temporal boundaries. Furthermore, we propose a simple yet effective method to capture the dependencies among different steps, which can be easily plugged into conventional proposal-based action detection methods for localizing important steps in instruction videos. In order to provide a benchmark for instruction video analysis, we evaluate plenty of approaches on the COIN dataset under different evaluation criteria. We expect the introduction of the COIN dataset will promote the future in-depth research on instruction video analysis for the community.

In this paper, we propose a generic Runtime Network Routing (RNR) framework for efficient image classification, which selects an optimal path inside the network. Unlike existing static neural network acceleration methods, our method preserves the full ability of the original large network and conducts dynamic routing at runtime according to the input image and current feature maps. The routing is performed in a bottom-up, layer-by-layer manner, where we model it as a Markov decision process and use reinforcement learning for training. The agent determines the estimated reward of each sub-path and conducts routing conditioned on different samples, where a faster path is taken when the image is easier for the task. Since the ability of network is fully preserved, the balance point is easily adjustable according to the available resources. We test our method on both multi-path residual networks and incremental convolutional channel pruning, and show that RNR consistently outperforms static methods at the same computation complexity on both the CIFAR and ImageNet datasets. Our method can also be applied to off-the-shelf neural network structures and easily extended to other application scenarios.

In this paper, we propose a simple yet effective method to learn globally optimized detector for object detection, which is a simple modification to the standard cross-entropy gradient inspired by the REINFORCE algorithm. In our approach, the cross-entropy gradient is adaptively adjusted according to overall mean Average Precision (mAP) of the current state for each detection candidate, which leads to more effective gradient and global optimization of detection results, and brings no computational overhead. Benefiting from more precise gradients produced by the global optimization method, our framework significantly improves state-of-the-art object detectors. Furthermore, since our method is based on scores and bounding boxes without modification on the architecture of object detector, it can be easily applied to off-the-shelf modern object detection frameworks.

In this paper, we propose an attention-aware deep reinforcement learning (ADRL) method for video face recognition, which aims to discard the misleading and confounding frames and find the focuses of attentions in face videos for person recognition. We formulate the process of finding the attentions of videos as a Markov decision process and train the attention model through a deep reinforcement learning framework without using extra labels. Unlike existing attention models, our method takes information from both the image space and the feature space as the input to make better use of face information that is discarded in the feature learning process. Besides, our approach is attention-aware, which seeks different attentions of videos for the recognition of different pairs of videos. Our approach achieves very competitive video face recognition performance on three widely used video face datasets.

In this paper, we propose a discriminative aggregation network (DAN) method for video face recognition, which aims to integrate information from video frames effectively and efficiently. Unlike existing aggregation methods, our method aggregates raw video frames directly instead of the features obtained by complex processing. By combining the idea of metric learning and adversarial learning, we learn an aggregation network that produces more discriminative synthesized images compared to raw input frames. Our framework reduces the number of frames to be processed and significantly speed up the recognition procedure. Furthermore, low-quality frames containing misleading information are filtered and denoised during the aggregation process, which makes our system more robust and discriminative. Experimental results show that our method can generate discriminative images from video clips and improve the overall recognition performance in both the speed and accuracy on three widely used datasets.

Intelligent transportation systems, e.g., Uber, have become an important tool for urban transportation. An important problem is k nearest neighbor (kNN) search on moving objects with road-network constraints, which, given moving objects on the road networks and a query, finds k nearest objects to the query location. Existing studies focus on either kNN search on static objects or continuous kNN search with Euclidean-distance constraints. The former cannot support dynamic updates of moving objects while the latter cannot support road networks. Since the objects are dynamically moving on the road networks, there are two main challenges. The first is how to index the moving objects on road networks and the second is how to find the k nearest moving objects. To address these challenges, in this paper we proposes a new index, V-Tree, which has two salient features. Firstly, it is a balanced search tree and can support efficient kNN search. Secondly, it can support dynamical updates of moving objects. To build a V-Tree, we iteratively partition the road network into sub-networks and build a tree structure on top of the sub-networks. Then we associate the moving objects on their nearest vertices in the V-Tree. When the location of an object is updated, we only need to update the tree nodes on the path from the corresponding leaf node to the root. We design a novel kNN search algorithm using V-Tree by pruning large numbers of irrelevant vertices in the road network. Experimental results on real datasets show that our method significantly outperforms baseline approaches by 2-3 orders of magnitude.

In this paper, we propose a Runtime Neural Pruning (RNP) framework which prunes the deep neural network dynamically at the runtime. Unlike existing neural pruning methods which produce a fixed pruned model for deployment, our method preserves the full ability of the original network and conducts pruning according to the input image and current feature maps adaptively. The pruning is performed in a bottom-up, layer-by-layer manner, which we model as a Markov decision process and use reinforcement learning for training. The agent judges the importance of each convolutional kernel and conducts channel-wise pruning conditioned on different samples, where the network is pruned more when the image is easier for the task. Since the ability of network is fully preserved, the balance point is easily adjustable according to the available resources. Our method can be applied to off-the-shelf network structures and reach a better tradeoff between speed and accuracy, especially with a large pruning rate.