We present Florence-VL, a new family of multimodal large language models (MLLMs) with enriched visual representations produced by Florence-2 [45], a generative vision foundation model. Unlike the widely used CLIP-style vision transformer [35] trained by contrastive learning, Florence-2 can capture different levels and aspects of visual features, which are more versatile to be adapted to diverse downstream tasks. We propose a novel feature-fusion architecture and an innovative training recipe that effectively integrates Florence-2’s visual features into pre-trained LLMs, such as Phi 3.5 and LLama 3. In particular, we propose "depth-breath fusion (DBFusion)" to fuse the visual features extracted from different depths and under multiple prompts. Our model training is composed of end-to-end pretraining of the whole model followed by finetuning of the projection layer and the LLM, on a carefully designed recipe of diverse open-source datasets that include high-quality image captions and instruction-tuning pairs. Our quantitative analysis and visualization of Florence-VL’s visual features show its advantages over popular vision encoders on vision-language alignment, where the enriched depth and breath play important roles. Florence-VL achieves significant improvements over existing state-of-the-art MLLMs across various multi-modal and vision-centric benchmarks covering general VQA, perception, hallucination, OCR, Chart, knowledge-intensive understanding, etc. To facilitate future research, our models and the complete training recipe are open-sourced. https://github.com/JiuhaiChen/Florence-VL

LLMs have demonstrated impressive performance across various language tasks. However, the strengths of LLMs can vary due to different architectures, model sizes, areas of training data, etc. Therefore, ensemble reasoning for the strengths of different LLM experts is critical to achieving consistent and satisfactory performance on diverse inputs across a wide range of tasks. However, existing LLM ensemble methods are either computationally intensive or incapable of leveraging complementary knowledge among LLM experts for various inputs. In this paper, we propose an efficient Dynamic Ensemble Reasoning paradigm, called DER to integrate the strengths of multiple LLM experts conditioned on dynamic inputs. Specifically, we model the LLM ensemble reasoning problem as a Markov Decision Process, wherein an agent sequentially takes inputs to request knowledge from an LLM candidate and passes the output to a subsequent LLM candidate. Moreover, we devise a reward function to train a DER-Agent to dynamically select an optimal answering route given the input questions, aiming to achieve the highest performance with as few computational resources as possible. Last, to fully transfer the expert knowledge from the prior LLMs, we develop a Knowledge Transfer Prompt that enables the subsequent LLM candidates to transfer complementary knowledge effectively. Experiments demonstrate that our method uses fewer computational resources to achieve better performance compared to state-of-the-art baselines. Code and appendix are available at https://github.com/Fhujinwu/DER.

We introduce Florence-2, a novel vision foundation model with a unified, prompt-based representation for various computer vision and vision-language tasks. While existing large vision models excel in transfer learning, they struggle to perform diverse tasks with simple instructions, a capability that implies handling the complexity of various spatial hierarchy and semantic granularity. Florence-2 was designed to take text-prompt as task instructions and generate desirable results in text forms, whether it be captioning, object detection, grounding or segmentation. This multi-task learning setup demands large-scale, high-quality annotated data. To this end, we co-developed FLD-5B that consists of 5.4 billion comprehensive visual annotations on 126 million images, using an iterative strategy of automated image annotation and model refinement. We adopted a sequence-to-sequence structure to train Florence-2 to perform versatile and comprehensive vision tasks. Extensive evaluations on numerous tasks demonstrated Florence-2 to be a strong vision foundation model contender with un-precedented zero-shot and fine-tuning capabilities.

In this work, we present efficient modulation, a novel design for efficient vision networks. We revisit the modulation mechanism, which operates input through convolutional context modeling and feature projection layers, and fuses features via element-wise multiplication and an MLP block. We demonstrate that the abstracted modulation mechanism is particularly well suited for efficient networks and further tailor the modulation design by proposing the efficient modulation (EfficientMod) block, which is considered the essential building block for our networks. Bene- fiting from the prominent representational ability of modulation mechanism and the efficiency of efficient modulation design, our network can accomplish better accuracy-efficiency trade-offs and set new state-of-the-art performance for efficient networks. When integrating EfficientMod block with the vanilla self-attention block, we obtain the hybrid architecture and further improve the performance without sacrificing the efficiency. We carry out comprehensive experiments to verify EfficientMod’s performance. With fewer parameters, our EfficientMod-s performs 0.6 top-1 accuracy better than the prior state-of-the-art approach EfficientFormerV2-s2 without any training tricks and is 25% faster on GPU. Additionally, our method presents a notable improvement in downstream tasks, outperforming EfficientFormerV2-s by 3.6 mIoU on the ADE20K benchmark. Code and checkpoints are available at https://github.com/ma-xu/EfficientMod.

Human intelligence is multimodal; we integrate visual, linguistic, and acoustic signals to maintain a holistic worldview. Most current pretraining methods, however, are limited to one or two modalities. We present i-Code, a self-supervised pretraining framework where users may flexibly combine the modalities of vision, speech, and language into unified and general-purpose vector representations. In this framework, data from each modality are first given to pretrained single-modality encoders. The encoder outputs are then integrated with a multimodal fusion network, which uses novel merge- and co-attention mechanisms to effectively combine information from the different modalities. The entire system is pretrained end-to-end with new objectives including masked modality unit modeling and cross-modality contrastive learning. Unlike previous research using only video for pretraining, the i-Code framework can dynamically process single, dual, and triple-modality data during training and inference, flexibly projecting different combinations of modalities into a single representation space. Experimental results demonstrate how i-Code can outperform state-of-the-art techniques on five multimodal understanding tasks and single-modality benchmarks, improving by as much as 11% and demonstrating the power of integrative multimodal pretraining.

In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers’ behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by 50%, while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by 1.4 - 7.8× compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of 41.1% on ImageNet, surpassing the original CLIP ViT-B/16 by 3.5% while utilizing only 8.9% parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip.

Visual recognition is recently learned via either super-vised learning on human-annotated image-label data or language-image contrastive learning with webly-crawled image-text pairs. While supervised learning may result in a more discriminative representation, language-image pretraining shows unprecedented zero-shot recognition ca-pability, largely due to the different properties of data sources and learning objectives. In this work, we intro-duce a new formulation by combining the two data sources into a common image-text-label space. In this space, we propose a new learning paradigm, called Unified Con-trastive Learning (UniCL) with a single learning objective to seamlessly prompt the synergy of two data types. Ex-tensive experiments show that our UniCL is an effective way of learning semantically rich yet discriminative repre-sentations, universally for image recognition in zero-shot, linear-probing, fully finetuning and transfer learning sce-narios. Particularly, it attains gains up to 9.2% and 14.5% in average on zero-shot recognition benchmarks over the language-image contrastive learning and supervised learning methods, respectively. In linear probe setting, it also boosts the performance over the two methods by 7.3% and 3.4%, respectively. Our study also indicates that UniCL stand-alone is a good learner on pure image-label data, rivaling the supervised learning methods across three im-age classification datasets and two types of vision back-bones, ResNet and Swin Transformer. Code is available at: https://github.com/microsoft/UniCL.

Vision Transformer (ViT) models have recently drawn much attention in computer vision due to their high model capability. However, ViT models suffer from huge number of parameters, restricting their applicability on devices with limited memory. To alleviate this problem, we propose MiniViT, a new compression framework, which achieves parameter reduction in vision transformers while retaining the same performance. The central idea of MiniViT is to multiplex the weights of consecutive transformer blocks. More specifically, we make the weights shared across layers, while imposing a transformation on the weights to increase diversity. Weight distillation over self-attention is also applied to transfer knowledge from large-scale ViT models to weight-multiplexed compact models. Comprehensive experiments demonstrate the efficacy of MiniViT, showing that it can reduce the size of the pre-trained Swin-B transformer by 48%, while achieving an increase of 1.0% in Top-1 accuracy on ImageNet. Moreover, using a single-layer of parameters, MiniViT is able to compress DeiT-B by 9.7 times from 86M to 9M parameters, without seriously compromising the performance. Finally, we verify the transferability of MiniViT by reporting its performance on downstream benchmarks. Code and models are available at here.

The complex nature of combining localization and classification in object detection has resulted in the flourished development of methods. Previous works tried to improve the performance in various object detection heads but failed to present a unified view. In this paper, we present a novel dynamic head framework to unify object detection heads with attentions. By coherently combining multiple self-attention mechanisms between feature levels for scale-awareness, among spatial locations for spatial-awareness, and within output channels for task-awareness, the proposed approach significantly improves the representation ability of object detection heads without any computational overhead. Further experiments demonstrate that the effectiveness and efficiency of the proposed dynamic head on the COCO benchmark. With a standard ResNeXt-101-DCN backbone, we largely improve the performance over popular object detectors and achieve a new state-of-the-art at 54.0 AP. The code will be released at https://github.com/microsoft/DynamicHead.

In this paper, we are interested in the bottom-up paradigm of estimating human poses from an image. We study the dense keypoint regression framework that is previously inferior to the keypoint detection and grouping framework. Our motivation is that regressing keypoint positions accurately needs to learn representations that focus on the keypoint regions.We present a simple yet effective approach, named disentangled keypoint regression (DEKR). We adopt adaptive convolutions through pixel-wise spatial transformer to activate the pixels in the keypoint regions and accordingly learn representations from them. We use a multi-branch structure for separate regression: each branch learns a representation with dedicated adaptive convolutions and regresses one keypoint. The resulting disentangled representations are able to attend to the keypoint regions, respectively, and thus the keypoint regression is spatially more accurate. We empirically show that the proposed direct regression method outperforms keypoint detection and grouping methods and achieves superior bottom-up pose estimation results on two benchmark datasets, COCO and CrowdPose. The code and models are available at https://github.com/HRNet/DEKR.

We present an efficient high-resolution network, Lite-HRNet, for human pose estimation. We start by simply applying the efficient shuffle block in ShuffleNet to HRNet (high-resolution network), yielding stronger performance over popular lightweight networks, such as MobileNet, ShuffleNet, and Small HRNet. We find that the heavily-used pointwise (1 × 1) convolutions in shuffle blocks become the computational bottleneck. We introduce a lightweight unit, conditional channel weighting, to replace costly pointwise (1 × 1) convolutions in shuffle blocks. The complexity of channel weighting is linear w.r.t the number of channels and lower than the quadratic time complexity for pointwise convolutions. Our solution learns the weights from all the channels and over multiple resolutions that are readily available in the parallel branches in HRNet. It uses the weights as the bridge to exchange information across channels and resolutions, compensating the role played by the pointwise (1 × 1) convolution. Lite-HRNet demonstrates superior results on human pose estimation over popular lightweight networks. Moreover, Lite-HRNet can be easily applied to semantic segmentation task in the same lightweight manner. The code and models have been publicly available at https://github.com/HRNet/Lite-HRNet.

We present in this paper a new architecture, named Convolutional vision Transformer (CvT), that improves Vision Transformer (ViT) in performance and efficiency by introducing convolutions into ViT to yield the best of both de-signs. This is accomplished through two primary modifications: a hierarchy of Transformers containing a new convolutional token embedding, and a convolutional Transformer block leveraging a convolutional projection. These changes introduce desirable properties of convolutional neural networks (CNNs) to the ViT architecture (i.e. shift, scale, and distortion invariance) while maintaining the merits of Transformers (i.e. dynamic attention, global context, and better generalization). We validate CvT by conducting extensive experiments, showing that this approach achieves state-of-the-art performance over other Vision Transformers and ResNets on ImageNet-1k, with fewer parameters and lower FLOPs. In addition, performance gains are maintained when pretrained on larger datasets (e.g. ImageNet-22k) and fine-tuned to downstream tasks. Pretrained on ImageNet-22k, our CvT-W24 obtains a top-1 accuracy of 87.7% on the ImageNet-1k val set. Finally, our results show that the positional encoding, a crucial component in existing Vision Transformers, can be safely re-moved in our model, simplifying the design for higher resolution vision tasks. Code will be released at https://github.com/microsoft/CvT.

This paper presents a new Vision Transformer (ViT) architecture Multi-Scale Vision Longformer, which significantly enhances the ViT of [12] for encoding high-resolution images using two techniques. The first is the multi-scale model structure, which provides image encodings at multiple scales with manageable computational cost. The second is the attention mechanism of Vision Long-former, which is a variant of Longformer [3], originally developed for natural language processing, and achieves a linear complexity w.r.t. the number of input tokens. A comprehensive empirical study shows that the new ViT significantly outperforms several strong baselines, including the existing ViT models and their ResNet counterparts, and the Pyramid Vision Transformer from a concurrent work [47], on a range of vision tasks, including image classification, object detection, and segmentation. The models and source code are released at https://github.com/microsoft/vision-longformer.

Recently, Vision Transformer and its variants have shown great promise on various computer vision tasks. The ability to capture local and global visual dependencies through self-attention is the key to its success. But it also brings challenges due to quadratic computational overhead, especially for the high-resolution vision tasks(e.g., object detection). Many recent works have attempted to reduce the cost and improve model performance by applying either coarse-grained global attention or fine-grained local attention. However, both approaches cripple the modeling power of the original self-attention mechanism of multi-layer Transformers, leading to sub-optimal solutions. In this paper, we present focal attention, a new attention mechanism that incorporates both fine-grained local and coarse-grained global interactions. In this new mechanism, each token attends its closest surrounding tokens at the fine granularity and the tokens far away at a coarse granularity and thus can capture both short- and long-range visual dependencies efficiently and effectively. With focal attention, we propose a new variant of Vision Transformer models, called Focal Transformers, which achieve superior performance over the state-of-the-art (SoTA) Vision Transformers on a range of public image classification and object detection benchmarks. In particular, our Focal Transformer models with a moderate size of 51.1M and a large size of 89.8M achieve 83.6% and 84.0%Top-1 accuracy, respectively, on ImageNet classification at 224×224. When employed as the backbones, Focal Transformers achieve consistent and substantial improvements over the current SoTA Swin Transformers [44] across 6 different object detection methods. Our largest Focal Transformer yields58.7/59.0boxmAPs and50.9/51.3mask mAPs on COCO mini-val/test-dev, and55.4mIoU onADE20K for semantic segmentation, creating new SoTA on three of the most challenging computer vision tasks.

High-resolution representations are essential for position-sensitive vision problems, such as human pose estimation, semantic segmentation, and object detection. Existing state-of-the-art frameworks first encode the input image as a low-resolution representation through a subnetwork that is formed by connecting high-to-low resolution convolutions in series (e.g., ResNet, VGGNet), and then recover the high-resolution representation from the encoded low-resolution representation. Instead, our proposed network, named as High-Resolution Network (HRNet), maintains high-resolution representations through the whole process. There are two key characteristics: (i) Connect the high-to-low resolution convolution streams in parallel and (ii) repeatedly exchange the information across resolutions. The benefit is that the resulting representation is semantically richer and spatially more precise. We show the superiority of the proposed HRNet in a wide range of applications, including human pose estimation, semantic segmentation, and object detection, suggesting that the HRNet is a stronger backbone for computer vision problems. All the codes are available at https://github.com/HRNet.

In this work, we tackle the problem of estimating 3D human pose in camera space from a monocular image. First, we propose to use densely-generated limb depth maps to ease the learning of body joints depth, which are well aligned with image cues. Then, we design a lifting module from 2D pixel coordinates to 3D camera coordinates which explicitly takes the depth values as inputs, and is aligned with camera perspective projection model. We show our method achieves superior performance on large-scale 3D pose datasets Human3.6M and MPI-INF-3DHP, and sets the new state-of-the-art.

Bottom-up human pose estimation methods have difficulties in predicting the correct pose for small persons due to challenges in scale variation. In this paper, we present HigherHRNet: a novel bottom-up human pose estimation method for learning scale-aware representations using high-resolution feature pyramids. Equipped with multi-resolution supervision for training and multi-resolution aggregation for inference, the proposed approach is able to solve the scale variation challenge in bottom-up multi-person pose estimation and localize keypoints more precisely, especially for small person. The feature pyramid in HigherHRNet consists of feature map outputs from HRNet and upsampled higher-resolution outputs through a transposed convolution. HigherHRNet outperforms the previous best bottom-up method by 2.5% AP for medium person on COCO test-dev, showing its effectiveness in handling scale variation. Furthermore, HigherHRNet achieves new state-of-the-art result on COCO test-dev (70.5% AP) without using refinement or other post-processing techniques, surpassing all existing bottom-up methods. HigherHRNet even surpasses all top-down methods on CrowdPose test (67.6% AP), suggesting its robustness in crowded scene.

In this paper, we are interested in the human pose estimation problem with a focus on learning reliable high-resolution representations. Most existing methods recover high-resolution representations from low-resolution representations produced by a high-to-low resolution network. Instead, our proposed network maintains high-resolution representations through the whole process. We start from a high-resolution subnetwork as the first stage, gradually add high-to-low resolution subnetworks one by one to form more stages, and connect the mutli-resolution subnetworks in parallel. We conduct repeated multi-scale fusions such that each of the high-to-low resolution representations receives information from other parallel representations over and over, leading to rich high-resolution representations. As a result, the predicted keypoint heatmap is potentially more accurate and spatially more precise. We empirically demonstrate the effectiveness of our network through the superior pose estimation results over two benchmark datasets: the COCO keypoint detection dataset and the MPII Human Pose dataset. In addition, we show the superiority of our network in pose tracking on the PoseTrack dataset. The code and models have been publicly available at https://github.com/leoxiaobin/deep-high-resolution-net.pytorch.

Deep learning gains lots of attentions in recent years and is more and more important for mining values in big data. However, to make deep learning practical for a wide range of applications in Tencent Inc., three requirements must be considered: 1) Lots of computational power are required to train a practical model with tens of millions of parameters and billions of samples for products such as automatic speech recognition (ASR), and the number of parameters and training data is still growing. 2) The capability of training larger model is necessary for better model quality. 3) Easy to use frameworks are valuable to do many experiments to perform model selection, such as finding an appropriate optimization algorithm and tuning optimal hyper-parameters. To accelerate training, support large models, and make experiments easier, we built Mariana, the Tencent deep learning platform, which utilizes GPU and CPU cluster to train models parallelly with three frameworks: 1) a multi-GPU data parallelism framework for deep neural networks (DNNs). 2) a multi-GPU model parallelism and data parallelism framework for deep convolutional neural networks (CNNs). 3) a CPU cluster framework for large scale DNNs. Mariana also provides built-in algorithms and features to facilitate experiments. Mariana is in production usage for more than one year, achieves state-of-the-art acceleration performance, and plays a key role in training models and improving quality for automatic speech recognition and image recognition in Tencent WeChat, a mobile social platform, and for Ad click-through rate prediction (pCTR) in Tencent QQ, an instant messaging platform, and Tencent Qzone, a social networking service.