ICCV 2021 论文列表

Efficient reasoning about the semantic, spatial, and temporal structure of a scene is a crucial prerequisite for autonomous driving. We present NEural ATtention fields (NEAT), a novel representation that enables such reasoning for end-to-end imitation learning models. NEAT is a continuous function which maps locations in Bird's Eye View (BEV) scene coordinates to waypoints and semantics, using intermediate attention maps to iteratively compress high-dimensional 2D image features into a compact representation. This allows our model to selectively attend to relevant regions in the input while ignoring information irrelevant to the driving task, effectively associating the images with the BEV representation. In a new evaluation setting involving adverse environmental conditions and challenging scenarios, NEAT outperforms several strong baselines and achieves driving scores on par with the privileged CARLA expert used to generate its training data. Furthermore, visualizing the attention maps for models with NEAT intermediate representations provides improved interpretability.

Fine-tuning from pre-trained ImageNet models has been a simple, effective, and popular approach for various computer vision tasks. The common practice of fine-tuning is to adopt a default hyperparameter setting with a fixed pre-trained model, while both of them are not optimized for specific tasks and time constraints. Moreover, in cloud computing or GPU clusters where the tasks arrive sequentially in a stream, faster online fine-tuning is a more desired and realistic strategy for saving money, energy consumption, and CO2 emission. In this paper, we propose a joint Neural Architecture Search and Online Adaption framework named NASOA towards a faster task-oriented fine-tuning upon the request of users. Specifically, NASOA first adopts an offline NAS to identify a group of training-efficient networks to form a pretrained model zoo. We propose a novel joint block and macro level search space to enable a flexible and efficient search. Then, by estimating fine-tuning performance via an adaptive model by accumulating experience from the past tasks, an online schedule generator is proposed to pick up the most suitable model and generate a personalized training regime with respect to each desired task in a one-shot fashion. The resulting model zoo is more training efficient than SOTA NAS models, e.g. 6x faster than RegNetY-16GF, and 1.7x faster than EfficientNetB3. Experiments on multiple datasets also show that NASOA achieves much better fine-tuning results, i.e. improving around 2.1% accuracy than the best performance in RegNet series under various time constraints and tasks; 40x faster compared to the BOHB method.

Recent advances on Out-of-Distribution (OoD) generalization reveal the robustness of deep learning models against distribution shifts. However, existing works focus on OoD algorithms, such as invariant risk minimization, domain generalization, or stable learning, without considering the influence of deep model architectures on OoD generalization, which may lead to sub-optimal performance. Neural Architecture Search (NAS) methods search for architecture based on its performance on the training data, which may result in poor generalization for OoD tasks. In this work, we propose robust Neural Architecture Search for OoD generalization (NAS-OoD), which optimizes the architecture with respect to its performance on generated OoD data by gradient descent. Specifically, a data generator is learned to synthesize OoD data by maximizing losses computed by different neural architectures, while the goal for architecture search is to find the optimal architecture parameters that minimize the synthetic OoD data losses. The data generator and the neural architecture are jointly optimized in an end-to-end manner, and the minimax training process effectively discovers robust architectures that generalize well for different distribution shifts. Extensive experimental results show that NAS-OoD achieves superior performance on various OoD generalization benchmarks with deep models having a much fewer number of parameters. In addition, on a real industry dataset, the proposed NAS-OoD method reduces the error rate by more than 70% compared with the state-of-the-art method, demonstrating the proposed method's practicality for real applications.

We introduce N-ImageNet, a large-scale dataset targeted for robust, fine-grained object recognition with event cameras. The dataset is collected using programmable hardware in which an event camera consistently moves around a monitor displaying images from ImageNet. N-ImageNet serves as a challenging benchmark for event-based object recognition, due to its large number of classes and samples. We empirically show that pretraining on N-ImageNet improves the performance of event-based classifiers and helps them learn with few labeled data. In addition, we present several variants of N-ImageNet to test the robustness of event-based classifiers under diverse camera trajectories and severe lighting conditions, and propose a novel event representation to alleviate the performance degradation. To the best of our knowledge, we are the first to quantitatively investigate the consequences caused by various environmental conditions on event-based object recognition algorithms. N-ImageNet and its variants are expected to guide practical implementations for deploying event-based object recognition algorithms in the real world.

Detection and segmentation of nuclei are fundamental analysis operations in pathology images, the assessments derived from which serve as the gold standard for cancer diagnosis. Manual segmenting nuclei is expensive and time-consuming. What's more, accurate segmentation detection of nuclei can be challenging due to the large appearance variation, conjoined and overlapping nuclei, and serious degeneration of histological structures. Supervised methods highly rely on massive annotated samples. The existing two unsupervised methods are prone to failure on degenerated samples. This paper proposes a Mutual-Complementing Framework (MCF) for nuclei detection and segmentation in pathology images. Two branches of MCF are trained in the mutual-complementing manner, where the detection branch complements the pseudo mask of the segmentation branch, while the progressive trained segmentation branch complements the missing nucleus templates through calculating the mask residual between the predicted mask and detected result. In the detection branch, two response map fusion strategies and gradient direction based postprocessing are devised to obtain the optimal detection response. Furthermore, the confidence loss combined with the synthetic samples and self-finetuning is adopted to train the segmentation network with only high confidence areas. Extensive experiments demonstrate that MCF achieves comparable performance with only a few nucleus patches as supervision. Especially, MCF possesses good robustness (only dropping by about 6%) on degenerated samples, which are critical and common cases in clinical diagnosis.

The classification and regression head are both indispensable components to build up a dense object detector, which are usually supervised by the same training samples and thus expected to have consistency with each other for detecting objects accurately in final detection pipelines. In this paper, we break the convention of the same training samples for these two heads in dense detectors and explore a novel supervisory paradigm, termed as Mutual Supervision (MuSu), to respectively and mutually assign training samples for the classification and regression head to ensure this consistency. MuSu defines training samples for the regression head mainly based on classification predicting scores and in turn, defines samples for the classification head based on localization scores from the regression head. Experimental results show that the convergence of detectors trained by this mutual supervision is guaranteed and the effectiveness of the proposed method is verified on the challenging MS COCO benchmark. We also find that tiling more anchors at the same location benefits detectors and leads to further improvements under this training scheme. We hope this work can inspire further researches on the interaction of the classification and regression task in detection and the supervision paradigm for detectors, especially separately for these two heads.

Existing blind image super-resolution (SR) methods mostly assume blur kernels are spatially invariant across the whole image. However, such an assumption is rarely applicable for real images whose blur kernels are usually spatially variant due to factors such as object motion and out-of-focus. Hence, existing blind SR methods would inevitably give rise to poor performance in real applications. To address this issue, this paper proposes a mutual affine network (MANet) for spatially variant kernel estimation. Specifically, MANet has two distinctive features. First, it has a moderate receptive field so as to keep the locality of degradation. Second, it involves a new mutual affine convolution (MAConv) layer that enhances feature expressiveness without increasing receptive field, model size and computation burden. This is made possible through exploiting channel interdependence, which applies each channel split with an affine transformation module whose input are the rest channel splits. Extensive experiments on synthetic and real images show that the proposed MANet not only performs favorably for both spatially variant and invariant kernel estimation, but also leads to state-of-the-art blind SR performance when combined with non-blind SR methods.

We present a multiview pseudo-labeling approach to video learning, a novel framework that uses complementary views in the form of appearance and motion information for semi-supervised learning in video. The complementary views help obtain more reliable "pseudo-labels"" on unlabeled video, to learn stronger video representations than from purely supervised data. Though our method capitalizes on multiple views, it nonetheless trains a model that is shared across appearance and motion input and thus, by design, incurs no additional computation overhead at inference time. On multiple video recognition datasets, our method substantially outperforms its supervised counterpart, and compares favorably to previous work on standard benchmarks in self-supervised video representation learning.

Dark environment becomes a challenge for computer vision algorithms owing to insufficient photons and undesirable noises. Most of the existing studies tackle this by either targeting human vision for better visual perception or improving the machine vision for specific high-level tasks. In addition, these methods rely on data argumentation and directly train their models based on real-world or over-simplified synthetic datasets without exploring the intrinsic pattern behind illumination translation. Here, we propose a novel multitask auto encoding transformation (MAET) model that combines human vision and machine vision tasks to enhance object detection in a dark environment. With a self-supervision learning, the MAET learns an intrinsic visual structure by encoding and decoding the realistic illumination-degrading transformation considering the physical noise model and image signal processing (ISP). Based on this representation, we achieve object detection task by decoding the bounding box coordinates and classes. To avoid the over-entanglement of two tasks, our MAET disentangles the object and degrading features by imposing an orthogonal tangent regularity. This forms a parametric manifold along which multitask predictions can be geometrically formulated by maximizing the orthogonality between the tangents along the outputs of respective tasks. Our framework can be implemented based on the mainstream object detection architecture and directly trained end-to-end using the normal target detection datasets, such as COCO and VOC. We have achieved the state-of-the-art performance using synthetic and real-world datasets.

This paper examines the problem of illumination spectra estimation in multispectral images. We cast the problem into a constrained matrix factorization problem and present a method for both single-global and multiple illumination estimation in which a deep unrolling network is constructed from the alternating direction method of multipliers(ADMM) optimization for solving the matrix factorization problem. To alleviate the lack of multispectral training data, we build a large multispectral reflectance image dataset for generating synthesized data and use them for training and evaluating our model. The results of simulations and real experiments demonstrate that the proposed method is able to outperform state-of-the-art spectral illumination estimation methods, and that it generalizes well to a wide variety of scenes and spectra.

We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10 more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast.

We introduce Multiresolution Deep Implicit Functions (MDIF), a hierarchical representation that can recover fine geometry detail, while being able to perform global operations such as shape completion. Our model represents a complex 3D shape with a hierarchy of latent grids, which can be decoded into different levels of detail and also achieve better accuracy. For shape completion, we propose latent grid dropout to simulate partial data in the latent space and therefore defer the completing functionality to the decoder side.This along with our multires design significantly improves the shape completion quality under decoder-only latent optimization. To the best of our knowledge, MDIF is the first deep implicit function model that can at the same time (1) represent different levels of detail and allow progressive decoding; (2) support both encoder-decoder inference and decoder-only latent optimization, and fulfill multiple applications; (3) perform detailed decoder-only shape completion. Experiments demonstrate its superior performance against prior art in various 3D reconstruction tasks.

In this paper, we investigate video summarization in the supervised setting. Since video summarization is subjective to the preference of the end-user, the design of a unique model is limited. In this work, we propose a model that provides personalized video summaries by conditioning the summarization process with predefined categorical user labels referred to as preferences. The underlying method is based on multiple pairwise rankers (called Multi-ranker), where the rankers are trained jointly to provide local summaries as well as a global summarization of a given video. In order to demonstrate the relevance and applications of our method in contrast with a classical global summarizer, we conduct experiments on multiple benchmark datasets, notably through a user study and comparisons with the state-of-art methods in the global video summarization task.

A few-shot font generation (FFG) method has to satisfy two objectives: the generated images should preserve the underlying global structure of the target character and present the diverse local reference style. Existing FFG methods aim to disentangle content and style either by extracting a universal representation style or extracting multiple component-wise style representations. However, previous methods either fail to capture diverse local styles or cannot be generalized to a character with unseen components, e.g., unseen language systems. To mitigate the issues, we propose a novel FFG method, named Multiple Localized Experts Few-shot Font Generation Network (MX-Font). MX-Font extracts multiple style features not explicitly conditioned on component labels, but automatically by multiple experts to represent different local concepts, e.g., left-side sub-glyph. Owing to the multiple experts, MX-Font can capture diverse local concepts and show the generalizability to unseen languages. During training, we utilize component labels as weak supervision to guide each expert to be specialized for different local concepts. We formulate the component assign problem to each expert as the graph matching problem, and solve it by the Hungarian algorithm. We also employ the independence loss and the content-style adversarial loss to impose the content-style disentanglement. In our experiments, MX-Font outperforms previous state-of-the-art FFG methods in the Chinese generation and cross-lingual, e.g., Chinese to Korean, generation.

The popularity of multimodal sensors and the accessibility of the Internet have brought us a massive amount of unlabeled multimodal data. Since existing datasets and well-trained models are primarily unimodal, the modality gap between a unimodal network and unlabeled multimodal data poses an interesting problem: how to transfer a pre-trained unimodal network to perform the same task on unlabeled multimodal data? In this work, we propose multimodal knowledge expansion (MKE), a knowledge distillation-based framework to effectively utilize multimodal data without requiring labels. Opposite to traditional knowledge distillation, where the student is designed to be lightweight and inferior to the teacher, we observe that the multimodal student model consistently rectifies pseudo labels and generalizes better than its teacher. Extensive experiments on four tasks and different modalities verify this finding. Furthermore, we connect the mechanism of MKE to semi-supervised learning and offer both empirical and theoretical explanations to understand the expansion capability of a multimodal student.

Survival outcome prediction is a challenging weakly-supervised and ordinal regression task in computational pathology that involves modeling complex interactions within the tumor microenvironment in gigapixel whole slide images (WSIs). Despite recent progress in formulating WSIs as bags for multiple instance learning (MIL), representation learning of entire WSIs remains an open and challenging problem, especially in overcoming: 1) the computational complexity of feature aggregation in large bags, and 2) the data heterogeneity gap in incorporating biological priors such as genomic measurements. In this work, we present a Multimodal Co-Attention Transformer (MCAT) framework that learns an interpretable, dense co-attention mapping between WSIs and genomic features formulated in an embedding space. Inspired by approaches in Visual Question Answering (VQA) that can attribute how word embeddings attend to salient objects in an image when answering a question, MCAT learns how histology patches attend to genes when predicting patient survival. In addition to visualizing multimodal interactions, our co-attention transformation also reduces the space complexity of WSI bags, which enables the adaptation of Transformer layers as a general encoder backbone in MIL. We apply our proposed method on five different cancer datasets (4,730 WSIs, 67 million patches). Our experimental results demonstrate that the proposed method consistently achieves superior performance compared to the state-of-the-art methods.

Multimodal self-supervised learning is getting more and more attention as it allows not only to train large networks without human supervision but also to search and retrieve data across various modalities. In this context, this paper proposes a framework that, starting from a pre-trained backbone, learns a common multimodal embedding space that, in addition to sharing representations across different modalities, enforces a grouping of semantically similar instances. To this end, we extend the concept of instance-level contrastive learning with a multimodal clustering step in the training pipeline to capture semantic similarities across modalities. The resulting embedding space enables retrieval of samples across all modalities, even from unseen datasets and different domains. To evaluate our approach, we train our model on the HowTo100M dataset and evaluate its zero-shot retrieval capabilities in two challenging domains, namely text-to-video retrieval, and temporal action localization, showing state-of-the-art results on four different datasets.

Spatio-temporal action detection is an important and challenging problem in video understanding. The existing action detection benchmarks are limited in aspects of small numbers of instances in a trimmed video or low-level atomic actions. This paper aims to present a new multi-person dataset of spatio-temporal localized sports actions, coined as MultiSports. We first analyze the important ingredients of constructing a realistic and challenging dataset for spatio-temporal action detection by proposing three criteria: (1) multi-person scenes and motion dependent identification, (2) with well-defined boundaries, (3) relatively fine-grained classes of high complexity. Based on these guidelines, we build the dataset of MultiSports v1.0 by selecting 4 sports classes, collecting 3200 video clips, and annotating 37701 action instances with 902k bounding boxes. Our datasets are characterized with important properties of high diversity, dense annotation, and high quality. Our MultiSports, with its realistic setting and detailed annotations, exposes the intrinsic challenges of spatio-temporal action detection. To benchmark this, we adapt several baseline methods to our dataset and give an in-depth analysis on the action detection results in our dataset. We hope our MultiSports can serve as a standard benchmark for spatio-temporal action detection in the future. Our dataset website is at https://deeperaction.github.io/multisports/.

Autonomous driving has attracted much attention over the years but turns out to be harder than expected, probably due to the difficulty of labeled data collection for model training. Self-supervised learning (SSL), which leverages unlabeled data only for representation learning, might be a promising way to improve model performance. Existing SSL methods, however, usually rely on the single-centric-object guarantee, which may not be applicable for multi-instance datasets such as street scenes. To alleviate this limitation, we raise two issues to solve: (1) how to define positive samples for cross-view consistency and (2) how to measure similarity in multi-instance circumstances. We first adopt an IoU threshold during random cropping to transfer global-inconsistency to local-consistency. Then, we propose two feature alignment methods to enable 2D feature maps for multi-instance similarity measurement. Additionally, we adopt intra-image clustering with self-attention for further mining intra-image similarity and translation-invariance. Experiments show that, when pre-trained on Waymo dataset, our method called Multi-instance Siamese Network (MultiSiam) remarkably improves generalization ability and achieves state-of-the-art transfer performance on autonomous driving benchmarks, including Cityscapes and BDD100K, while existing SSL counterparts like MoCo, MoCo-v2, and BYOL show significant performance drop. By pre-training on SODA10M, a large-scale autonomous driving dataset, MultiSiam exceeds the ImageNet pre-trained MoCo-v2, demonstrating the potential of domain-specific pre-training. Code will be available at https://github.com/KaiChen1998/MultiSiam .