The support set is a key to providing conditional prior for fast adaption of the model in few-shot tasks. But the strict form of support set makes its construction actually difficult in practical application. Motivated by ANIL, we rethink the role of adaption in the feature extractor of CNAPs, which is a state-of-the-art representative few-shot method. To investigate the role, Almost Zero-Shot (AZS) task is designed by fixing the support set to replace the common scheme, which provides corresponding support sets for the different conditional prior of different tasks. The AZS experiment results infer that the adaptation works little in the feature extractor. However, CNAPs cannot be robust to randomly selected support sets and perform poorly on some datasets of Meta-Dataset because of its scattered mean embeddings responded by the simple mean operator. To enhance the robustness of CNAPs, Canonical Mean Filter (CMF) module is proposed to make the mean embeddings intensive and stable in feature space by mapping the support sets into a canonical form. CMFs make CNAPs robust to any fixed support sets even if they are random matrices. This attribution makes CNAPs be able to remove the mean encoder and the parameter adaptation network at the test stage, while CNAP-CMF on AZS tasks keeps the performance with one-shot tasks. It leads to a big parameter reduction. Precisely, 40.48\% parameters are dropped at the test stage. Also, CNAP-CMF outperforms CNAPs in one-shot tasks because it addresses inner-task unstable performance problems. Classification performance, visualized and clustering results verify that CMFs make CNAPs better and simpler.
General point clouds have been increasingly investigated for different tasks, and recently Transformer-based networks are proposed for point cloud analysis. However, there are barely related works for medical point clouds, which are important for disease detection and treatment. In this work, we propose an attention-based model specifically for medical point clouds, namely 3D medical point Transformer (3DMedPT), to examine the complex biological structures. By augmenting contextual information and summarizing local responses at query, our attention module can capture both local context and global content feature interactions. However, the insufficient training samples of medical data may lead to poor feature learning, so we apply position embeddings to learn accurate local geometry and Multi-Graph Reasoning (MGR) to examine global knowledge propagation over channel graphs to enrich feature representations. Experiments conducted on IntrA dataset proves the superiority of 3DMedPT, where we achieve the best classification and segmentation results. Furthermore, the promising generalization ability of our method is validated on general 3D point cloud benchmarks: ModelNet40 and ShapeNetPart. Code is released.
We introduce PyTorchVideo, an open-source deep-learning library that provides a rich set of modular, efficient, and reproducible components for a variety of video understanding tasks, including classification, detection, self-supervised learning, and low-level processing. The library covers a full stack of video understanding tools including multimodal data loading, transformations, and models that reproduce state-of-the-art performance. PyTorchVideo further supports hardware acceleration that enables real-time inference on mobile devices. The library is based on PyTorch and can be used by any training framework; for example, PyTorchLightning, PySlowFast, or Classy Vision. PyTorchVideo is available at https://pytorchvideo.org/
Conventional video models rely on a single stream to capture the complex spatial-temporal features. Recent work on two-stream video models, such as SlowFast network and AssembleNet, prescribe separate streams to learn complementary features, and achieve stronger performance. However, manually designing both streams as well as the in-between fusion blocks is a daunting task, requiring to explore a tremendously large design space. Such manual exploration is time-consuming and often ends up with sub-optimal architectures when computational resources are limited and the exploration is insufficient. In this work, we present a pragmatic neural architecture search approach, which is able to search for two-stream video models in giant spaces efficiently. We design a multivariate search space, including 6 search variables to capture a wide variety of choices in designing two-stream models. Furthermore, we propose a progressive search procedure, by searching for the architecture of individual streams, fusion blocks, and attention blocks one after the other. We demonstrate two-stream models with significantly better performance can be automatically discovered in our design space. Our searched two-stream models, namely Auto-TSNet, consistently outperform other models on standard benchmarks. On Kinetics, compared with the SlowFast model, our Auto-TSNet-L model reduces FLOPS by nearly 11 times while achieving the same accuracy 78.9%. On Something-Something-V2, Auto-TSNet-M improves the accuracy by at least 2% over other methods which use less than 50 GFLOPS per video.
Automatic 3D neuron reconstruction is critical for analysing the morphology and functionality of neurons in brain circuit activities. However, the performance of existing tracing algorithms is hinged by the low image quality. Recently, a series of deep learning based segmentation methods have been proposed to improve the quality of raw 3D optical image stacks by removing noises and restoring neuronal structures from low-contrast background. Due to the variety of neuron morphology and the lack of large neuron datasets, most of current neuron segmentation models rely on introducing complex and specially-designed submodules to a base architecture with the aim of encoding better feature representations. Though successful, extra burden would be put on computation during inference. Therefore, rather than modifying the base network, we shift our focus to the dataset itself. The encoder-decoder backbone used in most neuron segmentation models attends only intra-volume voxel points to learn structural features of neurons but neglect the shared intrinsic semantic features of voxels belonging to the same category among different volumes, which is also important for expressive representation learning. Hence, to better utilise the scarce dataset, we propose to explicitly exploit such intrinsic features of voxels through a novel voxel-level cross-volume representation learning paradigm on the basis of an encoder-decoder segmentation model. Our method introduces no extra cost during inference. Evaluated on 42 3D neuron images from BigNeuron project, our proposed method is demonstrated to improve the learning ability of the original segmentation model and further enhancing the reconstruction performance.
Attention Mechanism is a widely used method for improving the performance of convolutional neural networks (CNNs) on computer vision tasks. Despite its pervasiveness, we have a poor understanding of what its effectiveness stems from. It is popularly believed that its effectiveness stems from the visual attention explanation, advocating focusing on the important part of input data rather than ingesting the entire input. In this paper, we find that there is only a weak consistency between the attention weights of features and their importance. Instead, we verify the crucial role of feature map multiplication in attention mechanism and uncover a fundamental impact of feature map multiplication on the learned landscapes of CNNs: with the high order non-linearity brought by the feature map multiplication, it played a regularization role on CNNs, which made them learn smoother and more stable landscapes near real samples compared to vanilla CNNs. This smoothness and stability induce a more predictive and stable behavior in-between real samples, and make CNNs generate better. Moreover, motivated by the proposed effectiveness of feature map multiplication, we design feature map multiplication network (FMMNet) by simply replacing the feature map addition in ResNet with feature map multiplication. FMMNet outperforms ResNet on various datasets, and this indicates that feature map multiplication plays a vital role in improving the performance even without finely designed attention mechanism in existing methods.
Current state-of-the-art object detection and segmentation methods work well under the closed-world assumption. This closed-world setting assumes that the list of object categories is available during training and deployment. However, many real-world applications require detecting or segmenting novel objects, i.e., object categories never seen during training. In this paper, we present, UVO (Unidentified Video Objects), a new benchmark for open-world class-agnostic object segmentation in videos. Besides shifting the problem focus to the open-world setup, UVO is significantly larger, providing approximately 8 times more videos compared with DAVIS, and 7 times more mask (instance) annotations per video compared with YouTube-VOS and YouTube-VIS. UVO is also more challenging as it includes many videos with crowded scenes and complex background motions. We demonstrated that UVO can be used for other applications, such as object tracking and super-voxel segmentation, besides open-world object segmentation. We believe that UVo is a versatile testbed for researchers to develop novel approaches for open-world class-agnostic object segmentation, and inspires new research directions towards a more comprehensive video understanding beyond classification and detection.
The standard way of training video models entails sampling at each iteration a single clip from a video and optimizing the clip prediction with respect to the video-level label. We argue that a single clip may not have enough temporal coverage to exhibit the label to recognize, since video datasets are often weakly labeled with categorical information but without dense temporal annotations. Furthermore, optimizing the model over brief clips impedes its ability to learn long-term temporal dependencies. To overcome these limitations, we introduce a collaborative memory mechanism that encodes information across multiple sampled clips of a video at each training iteration. This enables the learning of long-range dependencies beyond a single clip. We explore different design choices for the collaborative memory to ease the optimization difficulties. Our proposed framework is end-to-end trainable and significantly improves the accuracy of video classification at a negligible computational overhead. Through extensive experiments, we demonstrate that our framework generalizes to different video architectures and tasks, outperforming the state of the art on both action recognition (e.g., Kinetics-400 & 700, Charades, Something-Something-V1) and action detection (e.g., AVA v2.1 & v2.2).
We present a convolution-free approach to video classification built exclusively on self-attention over space and time. Our method, named "TimeSformer," adapts the standard Transformer architecture to video by enabling spatiotemporal feature learning directly from a sequence of frame-level patches. Our experimental study compares different self-attention schemes and suggests that "divided attention," where temporal attention and spatial attention are separately applied within each block, leads to the best video classification accuracy among the design choices considered. Despite the radically different design compared to the prominent paradigm of 3D convolutional architectures for video, TimeSformer achieves state-of-the-art results on several major action recognition benchmarks, including the best reported accuracy on Kinetics-400 and Kinetics-600. Furthermore, our model is faster to train and has higher test-time efficiency compared to competing architectures. Code and pretrained models will be made publicly available.