Data diversity and abundance are essential for improving the performance and generalization of models in natural language processing and 2D vision. However, the 3D vision domain suffers from a lack of 3D data, and simply combining multiple 3D datasets for pretraining a 3D backbone does not yield significant improvement, due to the domain discrepancies among different 3D datasets that impede effective feature learning. In this work, we identify the main sources of the domain discrepancies between 3D indoor scene datasets, and propose Swin3D++, an enhanced architecture based on Swin3D for efficient pretraining on multi-source 3D point clouds. Swin3D++ introduces domain-specific mechanisms to Swin3D's modules to address domain discrepancies and enhance the network capability on multi-source pretraining. Moreover, we devise a simple sourceaugmentation strategy to increase the pretraining data scale and facilitate supervised pretraining. We validate the effectiveness of our design, and demonstrate that Swin3D++ surpasses the state-of-the-art 3D pretraining methods on typical indoor scene understanding tasks.

The use of pretrained backbones with fine-tuning has shown success for 2D vision and natural language processing tasks, with advantages over task-specific networks. In this paper, we introduce a pretrained 3D backbone, called Swin3d, for 3D indoor scene understanding. We designed a 3D Swin Transformer as our backbone network, which enables efficient self-attention on sparse voxels with linear memory complexity, making the backbone scalable to large models and datasets. We also introduce a generalized contextual relative positional embedding scheme to capture various irregularities of point signals for improved network performance. We pretrained a large Swin3d model on a synthetic Structured3D dataset, which is an order of magnitude larger than the ScanNet dataset. Our model pretrained on the synthetic dataset not only generalizes well to downstream segmentation and detection on real 3D point datasets but also outperforms state-of-the-art methods on downstream tasks with +2.3 mIoU and +2.2 mIoU on S3DIS Area5 and 6-fold semantic segmentation, respectively, +1.8 mIoU on ScanNet segmentation (val), +1.9 [email protected] on ScanNet detection, and +8.1 [email protected] on S3DIS detection. A series of extensive ablation studies further validated the scalability, generality, and superior performance enabled by our approach.

Masked autoencoders (MAE) have recently been introduced to 3D self-supervised pretraining for point clouds due to their great success in NLP and computer vision. Unlike MAEs used in the image domain, where the pretext task is to restore features at the masked pixels, such as colors, the existing 3D MAE works reconstruct the missing geometry only, i.e, the location of the masked points. In contrast to previous studies, we advocate that point location recovery is inessential and restoring intrinsic point features is much superior. To this end, we propose to ignore point position reconstruction and recover high-order features at masked points including surface normals and surface variations, through a novel attention-based decoder which is independent of the encoder design. We validate the effectiveness of our pretext task and decoder design using different encoder structures for 3D training and demonstrate the advantages of our pretrained networks on various point cloud analysis tasks.

We introduce VASA, a framework for generating lifelike talking faces with appealing visual affective skills (VAS) given a single static image and a speech audio clip. Our premiere model, VASA-1, is capable of not only generating lip movements that are exquisitely synchronized with the audio, but also producing a large spectrum of facial nuances and natural head motions that contribute to the perception of authenticity and liveliness. The core innovations include a diffusion-based holistic facial dynamics and head movement generation model that works in a face latent space, and the development of such an expressive and disentangled face latent space using videos. Through extensive experiments including evaluation on a set of new metrics, we show that our method significantly outperforms previous methods along various dimensions comprehensively. Our method delivers high video quality with realistic facial and head dynamics and also supports the online generation of 512$\times$512 videos at up to 40 FPS with negligible starting latency. It paves the way for real-time engagements with lifelike avatars that emulate human conversational behaviors.

Abstract We present a learning based temporal supersampling algorithm for real‐time rendering. Different from existing learning‐based approaches that adopt an end‐to‐end training of a ‘black‐box’ neural network, we design a ‘white‐box’ solution that first classifies the pixels into different categories and then generates the supersampling result based on classification. Our key observation is that the core problem in temporal supersampling for rendering is to distinguish the pixels that consist of occlusion, aliasing, or shading changes. Samples from these pixels exhibit similar temporal radiance change but require different composition strategies to produce the correct supersampling result. Based on this observation, our method first classifies the pixels into several classes. Based on the classification results, our method then blends the current frame with the warped last frame via a learned weight map to get the supersampling results. We design compact neural networks for each step and develop dedicated loss functions for pixels belonging to different classes. Compared to existing learning based methods, our classifier‐based supersampling scheme takes less computational and memory cost for real‐time supersampling and generates visually compelling temporal supersampling results with fewer flickering artifacts. We evaluate the performance and generality of our method on several rendered game sequences and our method can upsample the rendered frames from 1080P to 2160P in just 13.39ms on a single Nvidia 3090GPU.

We present a method for creating 3D indoor scenes with a generative model learned from a collection of semantic-segmented depth images captured from different unknown scenes. Given a room with a specified size, our method automatically generates 3D objects in a room from a randomly sampled latent code. Different from existing methods that represent an indoor scene with the type, location, and other properties of objects in the room and learn the scene layout from a collection of complete 3D indoor scenes, our method models each indoor scene as a 3D semantic scene volume and learns a volumetric generative adversarial network (GAN) from a collection of 2.5D partial observations of 3D scenes. To this end, we apply a differentiable projection layer to project the generated 3D semantic scene volumes into semantic-segmented depth images and design a new multiple-view discriminator for learning the complete 3D scene volume from 2.5D semantic-segmented depth images. Compared to existing methods, our method not only efficiently reduces the workload of modeling and acquiring 3D scenes for training, but also produces better object shapes and their detailed layouts in the scene. We evaluate our method with different indoor scene datasets and demonstrate the advantages of our method. We also extend our method for generating 3D indoor scenes from semantic-segmented depth images inferred from RGB images of real scenes.1

Deep Convolutional Neural Networks (CNNs) have achieved remarkable success in various fields. However, training an excellent CNN is practically a trial-and-error process that consumes a tremendous amount of time and computer resources. To accelerate the training process and reduce the number of trials, experts need to understand what has occurred in the training process and why the resulting CNN behaves as it does. However, current popular training platforms, such as TensorFlow, only provide very little and general information, such as training/validation errors, which is far from enough to serve this purpose. To bridge this gap and help domain experts with their training tasks in a practical environment, we propose a visual analytics system, DeepTracker, to facilitate the exploration of the rich dynamics of CNN training processes and to identify the unusual patterns that are hidden behind the huge amount of information in training log. Specifically, we combine a hierarchical index mechanism and a set of hierarchical small multiples to help experts explore the entire training log from different levels of detail. We also introduce a novel cube-style visualization to reveal the complex correlations among multiple types of heterogeneous training data, including neuron weights, validation images, and training iterations. Three case studies are conducted to demonstrate how DeepTracker provides its users with valuable knowledge in an industry-level CNN training process; namely, in our case, training ResNet-50 on the ImageNet dataset. We show that our method can be easily applied to other state-of-the-art “very deep” CNN models.

Object skeleton detection is a challenging problem with wide application. Recently, deep Convolutional Neural Networks (CNNs) have substantially improved the performance of the state-of-the-art in this task. However, most of the existing CNN-Based methods are based on a skip-layer structure where low-level and high-level features are combined and learned so as to gather multi-level contextual information. As shallow features are too messy and lack semantic knowledge, they may cause errors and inaccuracy. Therefore, we propose a novel network architecture, Multi-Scale Bidirectional Fully Convolutional Network (MSB-FCN), to better capture and consolidate multi-scale high-level context information for object skeleton detection. Our network uses only deep features to build multi-scale feature representations, and employs a bidirectional structure to collect contextual knowledge. Hence the proposed MSB-FCN has the ability to learn the semantic-level information from different sub-regions. Furthermore, we introduce dense connections into the bidirectional structure of our MSB-FCN to ensure that the learning process at each scale can directly encode information from all other scales. Extensive experiments on various commonly used benchmarks demonstrate that the proposed MSB-FCN has achieved significant improvements over the state-of-the-art algorithms.

We introduce a View-Volume convolutional neural network (VVNet) for inferring the occupancy and semantic labels of a volumetric 3D scene from a single depth image. Our method extracts the detailed geometric features from the input depth image with a 2D view CNN and then projects the features into a 3D volume according to the input depth map via a projection layer. After that, we learn the 3D context information of the scene with a 3D volume CNN for computing the result volumetric occupancy and semantic labels. With combined 2D and 3D representations, the VVNet efficiently reduces the computational cost, enables feature extraction from multi-channel high resolution inputs, and thus significantly improve the result accuracy. We validate our method and demonstrate its efficiency and effectiveness on both synthetic SUNCG and real NYU dataset.

In this paper we present a novel network architecture, called Multi-Scale Cascade Network (MSC-Net), to identify the most visually conspicuous objects in an image. Our network consists of several stages (sub-networks) for handling saliency detection across different scales. All these sub-networks form a cascade structure (in a coarse-to-fine manner) where the same underlying convolutional feature representations are fully shared. Compared with existing CNN-based saliency models, the MSC-Net can naturally enable the learning process in the finer cascade stages to encode more global contextual information while progressively incorporating the saliency prior knowledge obtained from coarser stages and thus lead to better detection accuracy. We also design a novel refinement module to further filter out errors by considering the intermediate feedback information. Our MSC-Net is highly integrated, end-to-end trainable, and very powerful. The proposed method achieves state-of-the-art performance on five widely-used salient object detection benchmarks, outperforming existing methods and also maintaining high efficiency. Code and pre-trained models are available at https://github.com/lixin666/MSC-NET.

We present O-CNN , an Octree-based Convolutional Neural Network (CNN) for 3D shape analysis. Built upon the octree representation of 3D shapes, our method takes the average normal vectors of a 3D model sampled in the finest leaf octants as input and performs 3D CNN operations on the octants occupied by the 3D shape surface. We design a novel octree data structure to efficiently store the octant information and CNN features into the graphics memory and execute the entire O-CNN training and evaluation on the GPU. O-CNN supports various CNN structures and works for 3D shapes in different representations. By restraining the computations on the octants occupied by 3D surfaces, the memory and computational costs of the O-CNN grow quadratically as the depth of the octree increases, which makes the 3D CNN feasible for high-resolution 3D models. We compare the performance of the O-CNN with other existing 3D CNN solutions and demonstrate the efficiency and efficacy of O-CNN in three shape analysis tasks, including object classification, shape retrieval, and shape segmentation.