ICCV 2021 Papers

The domain-adaptive semantic segmentation in aerial images by a deep-learning technique remains a challenge owing to the domain gaps caused by a resolution, image sensors, time-zone, the density of buildings, and even building styles of each city. Currently, convolutional neural network (CNN)-based domain adaptation methodologies have been developed to decrease the domain gaps, but, they have shown still poor performance to utilize multiple aerial images in different domains. In this paper, therefore, the CNN-based network denoted as Self-Mutating Network, which changes the values of parameters of convolutional filters itself according to the domain of input image, is proposed. By adopting Parameter Mutation to change the values of parameters and Parameter Fluctuation to randomly convulse the parameters, the network self-changes and fine-tunes the parameters, then achieves better predictions of a domain-adaptive segmentation. Through the ablation study of the Self-Mutating Network, we concluded that the Self-Mutating Network can be utilized in the domain-adaptive semantic segmentation of aerial images in different domains.

Vision-Dialog Navigation (VDN) requires an agent to ask questions and navigate following the human responses to find target objects. Conventional approaches are only allowed to ask questions at predefined locations, which are built upon expensive dialogue annotations, and inconvenience the real-word human-robot communication and cooperation. In this paper, we propose a Self-Motivated Communication Agent (SCoA) that learns whether and what to communicate with human adaptively to acquire instructive information for realizing dialogue annotation-free navigation and enhancing the transferability in real-world unseen environment. Specifically, we introduce a whether-to-ask (WeTA) policy, together with uncertainty of which action to choose, to indicate whether the agent should ask a question. Then, a what-to-ask (WaTA) policy is proposed, in which, along with the oracle's answers, the agent learns to score question candidates so as to pick up the most informative one for navigation, and meanwhile mimic oracle's answering. Thus, the agent can navigate in a self-Q&A manner even in real-world environment where the human assistance is often unavailable. Through joint optimization of communication and navigation in a unified imitation learning and reinforcement learning framework, SCoA asks a question if necessary and obtains a hint for guiding the agent to move towards the target with less communication cost. Experiments on seen and unseen environments demonstrate that SCoA shows not only superior performance over existing baselines without dialog annotations, but also competing results compared with rich dialog annotations based counterparts.

The generalization capability of deep neural networks has been substantially improved by applying a wide spectrum of regularization methods, e.g., restricting function space, injecting randomness during training, augmenting data, etc. In this work, we propose a simple yet effective regularization method named progressive self-knowledge distillation (PS-KD), which progressively distills a model's own knowledge to soften hard targets (i.e., one-hot vectors) during training. Hence, it can be interpreted within a framework of knowledge distillation as a student becomes a teacher itself. Specifically, targets are adjusted adaptively by combining the ground-truth and past predictions from the model itself. We show that PS-KD provides an effect of hard example mining by rescaling gradients according to difficulty in classifying examples. The proposed method is applicable to any supervised learning tasks with hard targets and can be easily combined with existing regularization methods to further enhance the generalization performance. Furthermore, it is confirmed that PS-KD achieves not only better accuracy, but also provides high quality of confidence estimates in terms of calibration as well as ordinal ranking. Extensive experimental results on three different tasks, image classification, object detection, and machine translation, demonstrate that our method consistently improves the performance of the state-of-the-art baselines.

Bicubic downscaling is a prevalent technique used to reduce the video storage burden or to accelerate the downstream processing speed. However, the inverse upscaling step is non-trivial, and the downscaled video may also deteriorate the performance of downstream tasks. In this paper, we propose a self-conditioned probabilistic framework for video rescaling to learn the paired downscaling and upscaling procedures simultaneously. During the training, we decrease the entropy of the information lost in the downscaling by maximizing its probability conditioned on the strong spatial-temporal prior information within the downscaled video. After optimization, the downscaled video by our framework preserves more meaningful information, which is beneficial for both the upscaling step and the downstream tasks, e.g., video action recognition task. We further extend the framework to a lossy video compression system, in which a gradient estimator for non-differential industrial lossy codecs is proposed for the end-to-end training of the whole system. Extensive experimental results demonstrate the superiority and effectiveness of our approach on video rescaling, video compression, and efficient action recognition tasks.

In this work, we propose a camera self-calibration algorithm for generic cameras with arbitrary non-linear distortions. We jointly learn the geometry of the scene and the accurate camera parameters without any calibration objects. Our camera model consists of a pinhole model, a fourth order radial distortion, and a generic noise model that can learn arbitrary non-linear camera distortions. While traditional self-calibration algorithms mostly rely on geometric constraints, we additionally incorporate photometric consistency. This requires learning the geometry of the scene, and we use Neural Radiance Fields (NeRF). We also propose a new geometric loss function, viz., projected ray distance loss, to incorporate geometric consistency for complex non-linear camera models. We validate our approach on standard real image datasets and demonstrate that our model can learn the camera intrinsics and extrinsics (pose) from scratch without COLMAP initialization. Also, we show that learning accurate camera models in a differentiable manner allows us to improve PSNR over baselines.

Neural trees aim at integrating deep neural networks and decision trees so as to bring the best of the two worlds, including representation learning from the former and faster inference from the latter. In this paper, we introduce a novel approach, termed as Self-born Wiring (SeBoW), to learn neural trees from a mother deep neural network. In contrast to prior neural-tree approaches that either adopt a pre-defined structure or grow hierarchical layers in a progressive manner, task-adaptive neural trees in SeBoW evolve from a deep neural network through a construction-by-destruction process, enabling a global-level parameter optimization that further yields favorable results. Specifically, given a designated network configuration like VGG, SeBoW disconnects all the layers and derives isolated filter groups, based on which a global-level wiring process is conducted to attach a subset of filter groups, eventually bearing a lightweight neural tree. Extensive experiments demonstrate that, with a lower computational cost, SeBoW outperforms all prior neural trees by a significant margin and even achieves results on par with predominant non-tree networks like ResNets. Moreover, SeBoW proves its scalability to large-scale datasets like ImageNet, which has been barely explored by prior tree networks.

Deep learning has achieved remarkable progress for visual recognition on large-scale balanced datasets but still performs poorly on real-world long-tailed data. Previous methods often adopt class re-balanced training strategies to effectively alleviate the imbalance issue, but might be a risk of over-fitting tail classes. The recent decoupling method overcomes over-fitting issues by using a multi-stage training scheme, yet, it is still incapable of capturing tail class information in the feature learning stage. In this paper, we show that soft label can serve as a powerful solution to incorporate label correlation into a multi-stage training scheme for long-tailed recognition. The intrinsic relation between classes embodied by soft labels turns out to be helpful for long-tailed recognition by transferring knowledge from head to tail classes. Specifically, we propose a conceptually simple yet particularly effective multi-stage training scheme, termed as Self Supervised to Distillation (SSD). This scheme is composed of two parts. First, we introduce a self-distillation framework for long-tailed recognition, which can mine the label relation automatically. Second, we present a new distillation label generation module guided by self-supervision. The distilled labels integrate information from both label and data domains that can model long-tailed distribution effectively. We conduct extensive experiments and our method achieves the state-of-the-art results on three long-tailed recognition benchmarks: ImageNet-LT, CIFAR100-LT and iNaturalist 2018. Our SSD outperforms the strong LWS baseline by from 2.7% to 4.5% on various datasets.

Most action recognition solutions rely on dense sampling to precisely cover the informative temporal clip. Extensively searching temporal region is expensive for a real-world application. In this work, we focus on improving the inference efficiency of current action recognition backbones on trimmed videos, and illustrate that one action model can also cover then informative region by dropping non-informative features. We present Selective Feature Compression (SFC), an action recognition inference strategy that greatly increase model inference efficiency without any accuracy compromise. Differently from previous works that compress kernel sizes and decrease the channel dimension, we propose to compress feature flow at spatio-temporal dimension without changing any backbone parameters. Our experiments on Kinetics-400, UCF101 and ActivityNet show that SFC is able to reduce inference speed by 6-7x and memory usage by 5-6x compared with the commonly used 30 crops dense sampling procedure, while also slightly improving Top1 Accuracy. We thoroughly quantitatively and qualitatively evaluate SFC and all its components and show how does SFC learn to attend to important video regions and to drop temporal features that are uninformative for the task of action recognition.

Image segmentation is often ambiguous at the level of individual image patches and requires contextual information to reach label consensus. In this paper we introduce Segmenter, a transformer model for semantic segmentation. In contrast to convolution-based methods, our approach allows to model global context already at the first layer and throughout the network. We build on the recent Vision Transformer (ViT) and extend it to semantic segmentation. To do so, we rely on the output embeddings corresponding to image patches and obtain class labels from these embeddings with a point-wise linear decoder or a mask transformer decoder. We leverage models pre-trained for image classification and show that we can fine-tune them on moderate sized datasets available for semantic segmentation. The linear decoder allows to obtain excellent results already, but the performance can be further improved by a mask transformer generating class masks. We conduct an extensive ablation study to show the impact of the different parameters, in particular the performance is better for large models and small patch sizes. Segmenter attains excellent results for semantic segmentation. It outperforms the state of the art on both ADE20K and Pascal Context datasets and is competitive on Cityscapes.

Scene graph generation has emerged as an important problem in computer vision. While scene graphs provide a grounded representation of objects, their locations and relations in an image, they do so only at the granularity of proposal bounding boxes. In this work, we propose the first, to our knowledge, framework for pixel-level segmentation-grounded scene graph generation. Our framework is agnostic to the underlying scene graph generation method and address the lack of segmentation annotations in target scene graph datasets (e.g., Visual Genome) through transfer and multi-task learning from, and with, an auxiliary dataset (e.g., MS COCO). Specifically, each target object being detected is endowed with a segmentation mask, which is expressed as a lingual-similarity weighted linear combination over categories that have annotations present in an auxiliary dataset. These inferred masks, along with a Gaussian masking mechanism which grounds the relations at a pixel-level within the image, allow for improved relation prediction. The entire framework is end-to-end trainable and is learned in a multi-task manner.

In order to robustly deploy object detectors across a wide range of scenarios, they should be adaptable to shifts in the input distribution without the need to constantly annotate new data. This has motivated research in Unsupervised Domain Adaptation (UDA) algorithms for detection. UDA methods learn to adapt from labeled source domains to unlabeled target domains, by inducing alignment between detector features from source and target domains. Yet, there is no consensus on what features to align and how to do the alignment. In our work, we propose a framework that generalizes the different components commonly used by UDA methods laying the ground for an in-depth analysis of the UDA design space. Specifically, we propose a novel UDA algorithm, ViSGA, a direct implementation of our framework, that leverages the best design choices and introduces a simple but effective method to aggregate features at the instance-level based on the visual similarity before inducing group alignment via adversarial training. We show that both similarity-based grouping and adversarial training allows our model to focus on coarsely aligning feature groups, without being forced to match all instances across loosely aligned domains. Finally, we examine the applicability of ViSGA to the setting where labeled data are gathered from different sources. Experiments show that not only our method outperforms previous single-source approaches on Sim2Real and Adverse Weather, but also generalizes well to the multi-source setting.

Low-light video enhancement is an important task. Previous work is mostly trained on paired static images or videos. We compile a new dataset formed by our new strategy that contains high-quality spatially-aligned video pairs from dynamic scenes in low- and normal-light conditions. We built it using a mechatronic system to precisely control the dynamics during the video capture process, and further align the video pairs, both spatially and temporally, by identifying the system's uniform motion stage. Besides the dataset, we propose an end-to-end framework, in which we design a self-supervised strategy to reduce noise, while enhancing the illumination based on the Retinex theory. Extensive experiments based on various metrics and large-scale user study demonstrate the value of our dataset and effectiveness of our method. The dataset and code are available at https://github.com/dvlab-research/SDSD.

Remote sensing and automatic earth monitoring are key to solve global-scale challenges such as disaster prevention, land use monitoring, or tackling climate change. Although there exist vast amounts of remote sensing data, most of it remains unlabeled and thus inaccessible for supervised learning algorithms. Transfer learning approaches can reduce the data requirements of deep learning algorithms. However, most of these methods are pre-trained on ImageNet and their generalization to remote sensing imagery is not guaranteed due to the domain gap. In this work, we propose Seasonal Contrast (SeCo), an effective pipeline to leverage unlabeled data for in-domain pre-training of remote sensing representations. The SeCo pipeline is composed of two parts. First, a principled procedure to gather large-scale, unlabeled and uncurated remote sensing datasets containing images from multiple Earth locations at different timestamps. Second, a self-supervised algorithm that takes advantage of time and position invariance to learn transferable representations for remote sensing applications. We empirically show that models trained with SeCo achieve better performance than their ImageNet pre-trained counterparts and state-of-the-art self-supervised learning methods on multiple downstream tasks. The datasets and models in SeCo will be made public to facilitate transfer learning and enable rapid progress in remote sensing applications.

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.

We present a "learning to learn" approach for discovering white-box classification loss functions that are robust to label noise in the training data. We parameterise a flexible family of loss functions using Taylor polynomials, and apply evolutionary strategies to search for noise-robust losses in this space. To learn re-usable loss functions that can apply to new tasks, our fitness function scores their performance in aggregate across a range of training datasets and architectures. The resulting white-box loss provides a simple and fast "plug-and-play" module that enables effective label-noise-robust learning in diverse downstream tasks, without requiring a special training procedure or network architecture. The efficacy of our loss is demonstrated on a variety of datasets with both synthetic and real label noise, where we compare favourably to prior work.

We present a novel framework for controllable image restoration that can effectively restore multiple types and levels of degradation of a corrupted image. The proposed model, named TASNet, is automatically determined by our neural architecture search algorithm, which optimizes the efficiency-accuracy trade-off of the candidate model architectures. Specifically, we allow TASNet to share the early layers across different restoration tasks and adaptively adjust the remaining layers with respect to each task. The shared task-agnostic layers greatly improve the efficiency while the task-specific layers are optimized for restoration quality, and our search algorithm seeks for the best balance between the two. We also propose a new data sampling strategy to further improve the overall restoration performance. As a result, TASNet achieves significantly faster GPU latency and lower FLOPs compared to the existing state-of-the-art models, while also showing visually more pleasing outputs. The source code and pre-trained models are available at https://github.com/ghimhw/TASNet.

Light-field (LF) imaging is appealing to the mobile devices market because of its capability for intuitive post-capture processing. Acquiring LF data with high angular, spatial and temporal resolution poses significant challenges, especially with space constraints preventing bulky optics. At the same time, stereo video capture, now available on many consumer devices, can be interpreted as a sparse LF-capture. We explore the application of small baseline stereo videos for reconstructing high fidelity LF videos. We propose a self-supervised learning-based algorithm for LF video reconstruction from stereo video. The self-supervised LF video reconstruction is guided via the geometric information from the individual stereo pairs and the temporal information from the video sequence. LF estimation is further regularized by a low-rank constraint based on layered LF displays. The proposed self-supervised algorithm facilitates advantages such as post-training fine-tuning on test sequences and variable angular view interpolation and extrapolation. Quantitatively the LF videos show higher fidelity than previously proposed unsupervised approaches for LF reconstruction. We demonstrate our results via LF videos generated from stereo videos acquired from commercially available stereoscopic cameras. Finally, we demonstrate that our reconstructed LF videos allow applications such as post-capture focus control and RoI-based focus tracking for videos.

Scribble-supervised semantic segmentation has gained much attention recently for its promising performance without high-quality annotations. Due to the lack of supervision, confident and consistent predictions are usually hard to obtain. Typically, people handle these problems by either adopting an auxiliary task with the well-labeled dataset or incorporating a graphical model with additional requirements on scribble annotations. Instead, this work aims to achieve semantic segmentation by scribble annotations directly without extra information and other limitations. Specifically, we propose holistic operations, including minimizing entropy and a network embedded random walk on the neural representation to reduce uncertainty. Given the probabilistic transition matrix of a random walk, we further train the network with self-supervision on its neural eigenspace to impose consistency on predictions between related images. Comprehensive experiments and ablation studies verify the proposed approach, which demonstrates superiority over others; it is even comparable to some full-label supervised ones and works well when scribbles are randomly shrunk or dropped.

In this paper, we propose a progressive segmentation inference (PSI) framework to tackle with scribble-supervised semantic segmentation. In virtue of latent contextual dependency, we encapsulate two crucial cues, contextual pattern propagation and semantic label diffusion, to enhance and refine pixel-level segmentation results from partially known seeds. In contextual pattern propagation, different-granular contextual patterns are correlated and leveraged to properly diffuse pattern information based on graphical model, so as to increase the inference confidence of pixel label prediction. Further, depending on high confidence scores of estimated pixels, the initial annotated seeds are progressively spread over the image through dynamically learning an adaptive decision strategy. The two cues are finally modularized to form a close-looping update process during pixel-wise label inference. Extensive experiments demonstrate that our proposed progressive segmentation inference can benefit from the combination of spatial and semantic context cues, and meantime achieve the state-of-the-art performance on two public scribble segmentation datasets.