CVPR 2020 Papers

This paper proposes a new learning paradigm called filter grafting, which aims to improve the representation capability of Deep Neural Networks (DNNs). The motivation is that DNNs have unimportant (invalid) filters (e.g., l1 norm close to 0). These filters limit the potential of DNNs since they are identified as having little effect on the network. While filter pruning removes these invalid filters for efficiency consideration, filter grafting re-activates them from an accuracy boosting perspective. The activation is processed by grafting external information (weights) into invalid filters. To better perform the grafting process, we develop an entropy-based criterion to measure the information of filters and an adaptive weighting strategy for balancing the grafted information among networks. After the grafting operation, the network has very few invalid filters compared with its untouched state, enpowering the model with more representation capacity. We also perform extensive experiments on the classification and recognition tasks to show the superiority of our method. For example, the grafted MobileNetV2 outperforms the non-grafted MobileNetV2 by about 7 percent on CIFAR-100 dataset.

Difficulty in collecting and annotating large-scale video data raises a growing interest in learning models which can recognize novel classes with only a few training examples. In this paper, we propose the Ordered Temporal Alignment Module (OTAM), a novel few-shot learning framework that can learn to classify a previously unseen video. While most previous work neglects long-term temporal ordering information, our proposed model explicitly leverages the temporal ordering information in video data through ordered temporal alignment. This leads to strong data-efficiency for few-shot learning. In concrete, our proposed pipeline learns a deep distance measurement of the query video with respect to novel class proxies over its alignment path. We adopt an episode-based training scheme and directly optimize the few-shot learning objective. We evaluate OTAM on two challenging real-world datasets, Kinetics and Something-Something-V2, and show that our model leads to significant improvement of few-shot video classification over a wide range of competitive baselines and outperforms state-of-the-art benchmarks by a large margin.

Pill image recognition is vital for many personal/public health-care applications and should be robust to diverse unconstrained real-world conditions. Most existing pill recognition models are limited in tackling this challenging few-shot learning problem due to the insufficient instances per category. With limited training data, neural network-based models have limitations in discovering most discriminating features, or going deeper. Especially, existing models fail to handle the hard samples taken under less controlled imaging conditions. In this study, a new pill image database, namely CURE, is first developed with more varied imaging conditions and instances for each pill category. Secondly, a W2-net is proposed for better pill segmentation. Thirdly, a Multi-Stream (MS) deep network that captures task-related features along with a novel two-stage training methodology are proposed. Within the proposed framework, a Batch All strategy that considers all the samples is first employed for the sub-streams, and then a Batch Hard strategy that considers only the hard samples mined in the first stage is utilized for the fusion network. By doing so, complex samples that could not be represented by one type of feature could be focused and the model could be forced to exploit other domain-related information more effectively. Experiment results show that the proposed model outperforms state-of-the-art models on both the National Institute of Health (NIH) and our CURE database.

The problem of open-set recognition is considered. While previous approaches only consider this problem in the context of large-scale classifier training, we seek a unified solution for this and the low-shot classification setting. It is argued that the classic softmax classifier is a poor solution for open-set recognition, since it tends to overfit on the training classes. Randomization is then proposed as a solution to this problem. This suggests the use of meta-learning techniques, commonly used for few-shot classification, for the solution of open-set recognition. A new oPen sEt mEta LEaRning (PEELER) algorithm is then introduced. This combines the random selection of a set of novel classes per episode, a loss that maximizes the posterior entropy for examples of those classes, and a new metric learning formulation based on the Mahalanobis distance. Experimental results show that PEELER achieves state of the art open set recognition performance for both few-shot and large-scale recognition. On CIFAR and miniImageNet, it achieves substantial gains in seen/unseen class detection AUROC for a given seen-class classification accuracy.

Conventional methods for object detection typically require a substantial amount of training data and preparing such high-quality training data is very labor-intensive. In this paper, we propose a novel few-shot object detection network that aims at detecting objects of unseen categories with only a few annotated examples. Central to our method are our Attention-RPN, Multi-Relation Detector and Contrastive Training strategy, which exploit the similarity between the few shot support set and query set to detect novel objects while suppressing false detection in the background. To train our network, we contribute a new dataset that contains 1000 categories of various objects with high-quality annotations. To the best of our knowledge, this is one of the first datasets specifically designed for few-shot object detection. Once our few-shot network is trained, it can detect objects of unseen categories without further training or fine-tuning. Our method is general and has a wide range of potential applications. We produce a new state-of-the-art performance on different datasets in the few-shot setting. The dataset link is https://github.com/fanq15/Few-Shot-Object-Detection-Dataset.

Learning with limited data is a key challenge for visual recognition. Many few-shot learning methods address this challenge by learning an instance embedding function from seen classes and apply the function to instances from unseen classes with limited labels. This style of transfer learning is task-agnostic: the embedding function is not learned optimally discriminative with respect to the unseen classes, where discerning among them leads to the target task. In this paper, we propose a novel approach to adapt the instance embeddings to the target classification task with a set-to-set function, yielding embeddings that are task-specific and are discriminative. We empirically investigated various instantiations of such set-to-set functions and observed the Transformer is most effective --- as it naturally satisfies key properties of our desired model. We denote this model as FEAT (few-shot embedding adaptation w/ Transformer) and validate it on both the standard few-shot classification benchmark and four extended few-shot learning settings with essential use cases, i.e., cross-domain, transductive, generalized few-shot learning, and low-shot learning. It archived consistent improvements over baseline models as well as previous methods, and established the new state-of-the-art results on two benchmarks.

Recently, deep neural networks are introduced as supervised discriminative models for the learning of 3D point cloud segmentation. Most previous supervised methods require a large number of training data with human annotation part labels to guide the training process to ensure the model's generalization abilities on test data. In comparison, we propose a novel 3D shape segmentation method that requires few labeled data for training. Given an input 3D shape, the training of our model starts with identifying a similar 3D shape with part annotations from a mini-pool of shape templates (e.g. 10 shapes). With the selected template shape, a novel Coherent Point Transformer is proposed to fully leverage the power of a deep neural network to smoothly morph the template shape towards the input shape. Then, based on the transformed template shapes with part labels, a newly proposed Part-specific Density Estimator is developed to learn a continuous part-specific probability distribution function on the entire 3D space with a batch consistency regularization term. With the learned part-specific probability distribution, our model is able to predict the part labels of a new input 3D shape in an end-to-end manner. We demonstrate that our proposed method can achieve remarkable segmentation results on the ShapeNet dataset with few shots, compared to previous supervised learning approaches.

The ability to incrementally learn new classes is crucial to the development of real-world artificial intelligence systems. In this paper, we focus on a challenging but practical few-shot class-incremental learning (FSCIL) problem. FSCIL requires CNN models to incrementally learn new classes from very few labelled samples, without forgetting the previously learned ones. To address this problem, we represent the knowledge using a neural gas (NG) network, which can learn and preserve the topology of the feature manifold formed by different classes. On this basis, we propose the TOpology-Preserving knowledge InCrementer (TOPIC) framework. TOPIC mitigates the forgetting of the old classes by stabilizing NG's topology and improves the representation learning for few-shot new classes by growing and adapting NG to new training samples. Comprehensive experimental results demonstrate that our proposed method significantly outperforms other state-of-the-art class-incremental learning methods on CIFAR100, miniImageNet, and CUB200 datasets.

Deep neural network compression techniques such as pruning and weight tensor decomposition usually require fine-tuning to recover the prediction accuracy when the compression ratio is high. However, conventional fine-tuning suffers from the requirement of a large training set and the time-consuming training procedure. This paper proposes a novel solution for knowledge distillation from label-free few samples to realize both data efficiency and training/processing efficiency. We treat the original network as "teacher-net" and the compressed network as "student-net". A 1x1 convolution layer is added at the end of each layer block of the student-net, and we fit the block-level outputs of the student-net to the teacher-net by estimating the parameters of the added layers. We prove that the added layer can be merged without adding extra parameters and computation cost during inference. Experiments on multiple datasets and network architectures verify the method's effectiveness on student-nets obtained by various network pruning and weight decomposition methods. Our method can recover student-net's accuracy to the same level as conventional fine-tuning methods in minutes while using only 1% label-free data of the full training data.

Video interpolation aims to synthesize non-existent frames between two consecutive frames. Although existing optical flow based methods have achieved promising results, they still face great challenges in dealing with the interpolation of complicated dynamic scenes, which include occlusion, blur or abrupt brightness change. This is mainly because these cases may break the basic assumptions of the optical flow estimation (i.e. smoothness, consistency). In this work, we devised a novel structure-to-texture generation framework which splits the video interpolation task into two stages: structure-guided interpolation and texture refinement. In the first stage, deep structure-aware features are employed to predict feature flows from two consecutive frames to their intermediate result, and further generate the structure image of the intermediate frame. In the second stage, based on the generated coarse result, a Frame Texture Compensator is trained to fill in detailed textures. To the best of our knowledge, this is the first work that attempts to directly generate the intermediate frame through blending deep features. Experiments on both the benchmark datasets and challenging occlusion cases demonstrate the superiority of the proposed framework over the state-of-the-art methods. Codes are available on https://github.com/CM-BF/FeatureFlow.

We present a fast feature-metric point cloud registration framework, which enforces the optimisation of registration by minimising a feature-metric projection error without correspondences. The advantage of the feature-metric projection error is robust to noise, outliers and density difference in contrast to the geometric projection error. Besides, minimising the feature-metric projection error does not need to search the correspondences so that the optimisation speed is fast. The principle behind the proposed method is that the feature difference is smallest if point clouds are aligned very well. We train the proposed method in a semi-supervised or unsupervised approach, which requires limited or no registration label data. Experiments demonstrate our method obtains higher accuracy and robustness than the state-of-the-art methods. Besides, experimental results show that the proposed method can handle significant noise and density difference, and solve both same-source and cross-source point cloud registration.

In this paper, we propose a state-of-the-art video denoising algorithm based on a convolutional neural network architecture. Until recently, video denoising with neural networks had been a largely under explored domain, and existing methods could not compete with the performance of the best patch-based methods. The approach we introduce in this paper, called FastDVDnet, shows similar or better performance than other state-of-the-art competitors with significantly lower computing times. In contrast to other existing neural network denoisers, our algorithm exhibits several desirable properties such as fast runtimes, and the ability to handle a wide range of noise levels with a single network model. The characteristics of its architecture make it possible to avoid using a costly motion compensation stage while achieving excellent performance. The combination between its denoising performance and lower computational load makes this algorithm attractive for practical denoising applications. We compare our method with different state-of-art algorithms, both visually and with respect to objective quality metrics.

Almost all previous deep learning-based multi-view stereo (MVS) approaches focus on improving reconstruction quality. Besides quality, efficiency is also a desirable feature for MVS in real scenarios. Towards this end, this paper presents a Fast-MVSNet, a novel sparse-to-dense coarse-to-fine framework, for fast and accurate depth estimation in MVS. Specifically, in our Fast-MVSNet, we first construct a sparse cost volume for learning a sparse and high-resolution depth map. Then we leverage a small-scale convolutional neural network to encode the depth dependencies for pixels within a local region to densify the sparse high-resolution depth map. At last, a simple but efficient Gauss-Newton layer is proposed to further optimize the depth map. On one hand, the high-resolution depth map, the data-adaptive propagation method and the Gauss-Newton layer jointly guarantee the effectiveness of our method. On the other hand, all modules in our Fast-MVSNet are lightweight and thus guarantee the efficiency of our approach. Besides, our approach is also memory-friendly because of the sparse depth representation. Extensive experimental results show that our method is 5 times and 14 times faster than Point-MVSNet and R-MVSNet, respectively, while achieving comparable or even better results on the challenging Tanks and Temples dataset as well as the DTU dataset. Code is available at https://github.com/svip-lab/FastMVSNet.

The reconstruction of shredded documents consists in arranging the pieces of paper (shreds) in order to reassemble the original aspect of such documents. This task is particularly relevant for supporting forensic investigation as documents may contain criminal evidence. As an alternative to the laborious and time-consuming manual process, several researchers have been investigating ways to perform automatic digital reconstruction. A central problem in automatic reconstruction of shredded documents is the pairwise compatibility evaluation of the shreds, notably for binary text documents. In this context, deep learning has enabled great progress for accurate reconstructions in the domain of mechanically-shredded documents. A sensitive issue, however, is that current deep model solutions require an inference whenever a pair of shreds has to be evaluated. This work proposes a scalable deep learning approach for measuring pairwise compatibility in which the number of inferences scales linearly (rather than quadratically) with the number of shreds. Instead of predicting compatibility directly, deep models are leveraged to asymmetrically project the raw shred content onto a common metric space in which distance is proportional to the compatibility. Experimental results show that our method has accuracy comparable to the state-of-the-art with a speed-up of about 22 times for a test instance with 505 shreds (20 mixed shredded-pages from different documents).

Significant progress has been made in Video Object Segmentation (VOS), the video object tracking task in its finest level. While the VOS task can be naturally decoupled into image semantic segmentation and video object tracking, significantly much more research effort has been made in segmentation than tracking. In this paper, we introduce "tracking-by-detection" into VOS which can coherently integrates segmentation into tracking, by proposing a new temporal aggregation network and a novel dynamic time-evolving template matching mechanism to achieve significantly improved performance. Notably, our method is entirely online and thus suitable for one-shot learning, and our end-to-end trainable model allows multiple object segmentation in one forward pass. We achieve new state-of-the-art performance on the DAVIS benchmark without complicated bells and whistles in both speed and accuracy, with a speed of 0.14 second per frame and J &F measure of 75.9% respectively.

Texture synthesis using deep neural networks can generate high quality and diversified textures. However, it usually requires a heavy optimization process. The following works accelerate the process by using feed-forward networks, but at the cost of scalability. diversity or quality. We propose a new efficient method that aims to simulate the optimization process while retains most of the properties. Our method takes a noise image and the gradients from a descriptor network as inputs, and synthesize a refined image with respect to the target image. The proposed method can synthesize images with better quality and diversity than the other fast synthesis methods do. Moreover, our method trained on a large scale dataset can generalize to synthesize unseen textures.

In this paper, the main task we aim to tackle is the multi-instance semi-supervised video object segmentation across a sequence of frames where only the first-frame box-level ground-truth is provided. Detection-based algorithms are widely adopted to handle this task, and the challenges lie in the selection of the matching method to predict the result as well as to decide whether to update the target template using the newly predicted result. The existing methods, however, make these selections in a rough and inflexible way, compromising their performance. To overcome this limitation, we propose a novel approach which utilizes reinforcement learning to make these two decisions at the same time. Specifically, the reinforcement learning agent learns to decide whether to update the target template according to the quality of the predicted result. The choice of the matching method will be determined at the same time, based on the action history of the reinforcement learning agent. Experiments show that our method is almost 10 times faster than the previous state-of-the-art method with even higher accuracy (region similarity of 69.1% on DAVIS 2017 dataset).

Diffeomorphic deformable image registration is crucial in many medical image studies, as it offers unique, special features including topology preservation and invertibility of the transformation. Recent deep learning-based deformable image registration methods achieve fast image registration by leveraging a convolutional neural network (CNN) to learn the spatial transformation from the synthetic ground truth or the similarity metric. However, these approaches often ignore the topology preservation of the transformation and the smoothness of the transformation which is enforced by a global smoothing energy function alone. Moreover, deep learning-based approaches often estimate the displacement field directly, which cannot guarantee the existence of the inverse transformation. In this paper, we present a novel, efficient unsupervised symmetric image registration method which maximizes the similarity between images within the space of diffeomorphic maps and estimates both forward and inverse transformations simultaneously. We evaluate our method on 3D image registration with a large scale brain image dataset. Our method achieves state-of-the-art registration accuracy and running time while maintaining desirable diffeomorphic properties.

Historically, the pursuit of efficient inference has been one of the driving forces behind the research into new deep learning architectures and building blocks. Some of the recent examples include: the squeeze-and-excitation module, depthwise separable convolutions in Xception, and the inverted bottleneck in MobileNet v2. Notably, in all of these cases, the resulting building blocks enabled not only higher efficiency, but also higher accuracy, and found wide adoption in the field. In this work, we further expand the arsenal of efficient building blocks for neural network architectures; but instead of combining standard primitives (such as convolution), we advocate for the replacement of these dense primitives with their sparse counterparts. While the idea of using sparsity to decrease the parameter count is not new, the conventional wisdom is that this reduction in theoretical FLOPs does not translate into real-world efficiency gains. We aim to correct this misconception by introducing a family of efficient sparse kernels for several hardware platforms, which we plan to open source for the benefit of the community. Equipped with our efficient implementation of sparse primitives, we show that sparse versions of MobileNet v1 and MobileNet v2 architectures substantially outperform strong dense baselines on the efficiency-accuracy curve. On Snapdragon 835 our sparse networks outperform their dense equivalents by 1.3 - 2.4x - equivalent to approximately one entire generation of improvement. We hope that our findings will facilitate wider adoption of sparsity as a tool for creating efficient and accurate deep learning architectures.