Masked Autoencoders (MAE) have shown great potentials in self-supervised pre-training for language and 2D image transformers. However, it still remains an open question on how to exploit masked autoencoding for learning 3D representations of irregular point clouds. In this paper, we propose Point-M2AE, a strong Multi-scale MAE pre-training framework for hierarchical self-supervised learning of 3D point clouds. Unlike the standard transformer in MAE, we modify the encoder and decoder into pyramid architectures to progressively model spatial geometries and capture both fine-grained and high-level semantics of 3D shapes. For the encoder that downsamples point tokens by stages, we design a multi-scale masking strategy to generate consistent visible regions across scales, and adopt a local spatial self-attention mechanism to focus on neighboring patterns. By multi-scale token propagation, the lightweight decoder gradually upsamples point tokens with complementary skip connections from the encoder, which further promotes the reconstruction from a global-to-local perspective. Extensive experiments demonstrate the state-of-the-art performance of Point-M2AE for 3D representation learning. With a frozen encoder after pre-training, Point-M2AE achieves 92.9% accuracy for linear SVM on ModelNet40, even surpassing some fully trained methods. By fine-tuning on downstream tasks, Point-M2AE achieves 86.43% accuracy on ScanObjectNN, +3.36% to the second-best, and largely benefits the few-shot classification, part segmentation and 3D object detection with the hierarchical pre-training scheme. Code will be available at https://github.com/ZrrSkywalker/Point-M2AE.
In this study, we propose Mixed and Masked Image Modeling (MixMIM), a simple but efficient MIM method that is applicable to various hierarchical Vision Transformers. Existing MIM methods replace a random subset of input tokens with a special MASK symbol and aim at reconstructing original image tokens from the corrupted image. However, we find that using the MASK symbol greatly slows down the training and causes training-finetuning inconsistency, due to the large masking ratio (e.g., 40% in BEiT). In contrast, we replace the masked tokens of one image with visible tokens of another image, i.e., creating a mixed image. We then conduct dual reconstruction to reconstruct the original two images from the mixed input, which significantly improves efficiency. While MixMIM can be applied to various architectures, this paper explores a simpler but stronger hierarchical Transformer, and scales with MixMIM-B, -L, and -H. Empirical results demonstrate that MixMIM can learn high-quality visual representations efficiently. Notably, MixMIM-B with 88M parameters achieves 85.1% top-1 accuracy on ImageNet-1K by pretraining for 600 epochs, setting a new record for neural networks with comparable model sizes (e.g., ViT-B) among MIM methods. Besides, its transferring performances on the other 6 datasets show MixMIM has better FLOPs / performance tradeoff than previous MIM methods. Code is available at https://github.com/Sense-X/MixMIM.
Vision Transformers (ViT) become widely-adopted architectures for various vision tasks. Masked auto-encoding for feature pretraining and multi-scale hybrid convolution-transformer architectures can further unleash the potentials of ViT, leading to state-of-the-art performances on image classification, detection and semantic segmentation. In this paper, our ConvMAE framework demonstrates that multi-scale hybrid convolution-transformer can learn more discriminative representations via the mask auto-encoding scheme. However, directly using the original masking strategy leads to the heavy computational cost and pretraining-finetuning discrepancy. To tackle the issue, we adopt the masked convolution to prevent information leakage in the convolution blocks. A simple block-wise masking strategy is proposed to ensure computational efficiency. We also propose to more directly supervise the multi-scale features of the encoder to boost multi-scale features. Based on our pretrained ConvMAE models, ConvMAE-Base improves ImageNet-1K finetuning accuracy by 1.4% compared with MAE-Base. On object detection, ConvMAE-Base finetuned for only 25 epochs surpasses MAE-Base fined-tuned for 100 epochs by 2.9% box AP and 2.2% mask AP respectively. Code and pretrained models are available at https://github.com/Alpha-VL/ConvMAE.
Accurate and reliable 3D detection is vital for many applications including autonomous driving vehicles and service robots. In this paper, we present a flexible and high-performance 3D detection framework, named MPPNet, for 3D temporal object detection with point cloud sequences. We propose a novel three-hierarchy framework with proxy points for multi-frame feature encoding and interactions to achieve better detection. The three hierarchies conduct per-frame feature encoding, short-clip feature fusion, and whole-sequence feature aggregation, respectively. To enable processing long-sequence point clouds with reasonable computational resources, intra-group feature mixing and inter-group feature attention are proposed to form the second and third feature encoding hierarchies, which are recurrently applied for aggregating multi-frame trajectory features. The proxy points not only act as consistent object representations for each frame, but also serve as the courier to facilitate feature interaction between frames. The experiments on largeWaymo Open dataset show that our approach outperforms state-of-the-art methods with large margins when applied to both short (e.g., 4-frame) and long (e.g., 16-frame) point cloud sequences. Specifically, MPPNet achieves 74.21%, 74.62% and 73.31% for vehicle, pedestrian and cyclist classes on the LEVEL 2 mAPH metric with 16-frame input.
With the growing popularity of smartphones, capturing high-quality images is of vital importance to smartphones. The cameras of smartphones have small apertures and small sensor cells, which lead to the noisy images in low light environment. Denoising based on a burst of multiple frames generally outperforms single frame denoising but with the larger compututional cost. In this paper, we propose an efficient yet effective burst denoising system. We adopt a three-stage design: noise prior integration, multi-frame alignment and multi-frame denoising. First, we integrate noise prior by pre-processing raw signals into a variance-stabilization space, which allows using a small-scale network to achieve competitive performance. Second, we observe that it is essential to adopt an explicit alignment for burst denoising, but it is not necessary to integrate a learning-based method to perform multi-frame alignment. Instead, we resort to a conventional and efficient alignment method and combine it with our multi-frame denoising network. At last, we propose a denoising strategy that processes multiple frames sequentially. Sequential denoising avoids filtering a large number of frames by decomposing multiple frames denoising into several efficient sub-network denoising. As for each sub-network, we propose an efficient multi-frequency denoising network to remove noise of different frequencies. Our three-stage design is efficient and shows strong performance on burst denoising. Experiments on synthetic and real raw datasets demonstrate that our method outperforms state-of-the-art methods, with less computational cost. Furthermore, the low complexity and high-quality performance make deployment on smartphones possible.
Self-attention based models such as vision transformers (ViTs) have emerged as a very competitive architecture alternative to convolutional neural networks (CNNs) in computer vision. Despite increasingly stronger variants with ever-higher recognition accuracies, due to the quadratic complexity of self-attention, existing ViTs are typically demanding in computation and model size. Although several successful design choices (e.g., the convolutions and hierarchical multi-stage structure) of prior CNNs have been reintroduced into recent ViTs, they are still not sufficient to meet the limited resource requirements of mobile devices. This motivates a very recent attempt to develop light ViTs based on the state-of-the-art MobileNet-v2, but still leaves a performance gap behind. In this work, pushing further along this under-studied direction we introduce EdgeViTs, a new family of light-weight ViTs that, for the first time, enable attention-based vision models to compete with the best light-weight CNNs in the tradeoff between accuracy and on-device efficiency. This is realized by introducing a highly cost-effective local-global-local (LGL) information exchange bottleneck based on optimal integration of self-attention and convolutions. For device-dedicated evaluation, rather than relying on inaccurate proxies like the number of FLOPs or parameters, we adopt a practical approach of focusing directly on on-device latency and, for the first time, energy efficiency. Specifically, we show that our models are Pareto-optimal when both accuracy-latency and accuracy-energy trade-offs are considered, achieving strict dominance over other ViTs in almost all cases and competing with the most efficient CNNs.
This work targets at using a general deep learning framework to synthesize free-viewpoint images of arbitrary human performers, only requiring a sparse number of camera views as inputs and skirting per-case fine-tuning. The large variation of geometry and appearance, caused by articulated body poses, shapes and clothing types, are the key bottlenecks of this task. To overcome these challenges, we present a simple yet powerful framework, named Generalizable Neural Performer (GNR), that learns a generalizable and robust neural body representation over various geometry and appearance. Specifically, we compress the light fields for novel view human rendering as conditional implicit neural radiance fields from both geometry and appearance aspects. We first introduce an Implicit Geometric Body Embedding strategy to enhance the robustness based on both parametric 3D human body model and multi-view images hints. We further propose a Screen-Space Occlusion-Aware Appearance Blending technique to preserve the high-quality appearance, through interpolating source view appearance to the radiance fields with a relax but approximate geometric guidance. To evaluate our method, we present our ongoing effort of constructing a dataset with remarkable complexity and diversity. The dataset GeneBody-1.0, includes over 360M frames of 370 subjects under multi-view cameras capturing, performing a large variety of pose actions, along with diverse body shapes, clothing, accessories and hairdos. Experiments on GeneBody-1.0 and ZJU-Mocap show better robustness of our methods than recent state-of-the-art generalizable methods among all cross-dataset, unseen subjects and unseen poses settings. We also demonstrate the competitiveness of our model compared with cutting-edge case-specific ones. Dataset, code and model will be made publicly available.
In this work, we tackle the problem of real-world fluid animation from a still image. The key of our system is a surface-based layered representation deriving from video decomposition, where the scene is decoupled into a surface fluid layer and an impervious background layer with corresponding transparencies to characterize the composition of the two layers. The animated video can be produced by warping only the surface fluid layer according to the estimation of fluid motions and recombining it with the background. In addition, we introduce surface-only fluid simulation, a $2.5D$ fluid calculation version, as a replacement for motion estimation. Specifically, we leverage the triangular mesh based on a monocular depth estimator to represent the fluid surface layer and simulate the motion in the physics-based framework with the inspiration of the classic theory of the hybrid Lagrangian-Eulerian method, along with a learnable network so as to adapt to complex real-world image textures. We demonstrate the effectiveness of the proposed system through comparison with existing methods in both standard objective metrics and subjective ranking scores. Extensive experiments not only indicate our method's competitive performance for common fluid scenes but also better robustness and reasonability under complex transparent fluid scenarios. Moreover, as the proposed surface-based layer representation and surface-only fluid simulation naturally disentangle the scene, interactive editing such as adding objects to the river and texture replacing could be easily achieved with realistic results.
6-DoF object pose estimation from a monocular image is challenging, and a post-refinement procedure is generally needed for high-precision estimation. In this paper, we propose a framework based on a recurrent neural network (RNN) for object pose refinement, which is robust to erroneous initial poses and occlusions. During the recurrent iterations, object pose refinement is formulated as a non-linear least squares problem based on the estimated correspondence field (between a rendered image and the observed image). The problem is then solved by a differentiable Levenberg-Marquardt (LM) algorithm enabling end-to-end training. The correspondence field estimation and pose refinement are conducted alternatively in each iteration to recover the object poses. Furthermore, to improve the robustness to occlusion, we introduce a consistency-check mechanism based on the learned descriptors of the 3D model and observed 2D images, which downweights the unreliable correspondences during pose optimization. Extensive experiments on LINEMOD, Occlusion-LINEMOD, and YCB-Video datasets validate the effectiveness of our method and demonstrate state-of-the-art performance.