ICCV 2021 论文列表

The recently proposed Detection Transformer (DETR) model successfully applies Transformer to objects detection and achieves comparable performance with two-stage object detection frameworks, such as Faster-RCNN. However, DETR suffers from its slow convergence. Training DETR from scratch needs 500 epochs to achieve a high accuracy. To accelerate its convergence, we propose a simple yet effective scheme for improving the DETR framework, namely Spatially Modulated Co-Attention (SMCA) mechanism. The core idea of SMCA is to conduct location-aware co-attention in DETR by constraining co-attention responses to be high near initially estimated bounding box locations. Our proposed SMCA increases DETR's convergence speed by replacing the original co-attention mechanism in the decoder while keeping other operations in DETR unchanged. Furthermore, by integrating multi-head and scale-selection attention designs into SMCA, our fully-fledged SMCA can achieve better performance compared to DETR with a dilated convolution-based backbone (45.6 mAP at 108 epochs vs. 43.3 mAP at 500 epochs). We perform extensive ablation studies on COCO dataset to validate SMCA. Code is released at https://github.com/gaopengcuhk/SMCA-DETR.

Virtual try-on tasks have drawn increased attention. Prior arts focus on tackling this task via warping clothes and fusing the information at the pixel level with the help of semantic segmentation. However, conducting semantic segmentation is time-consuming and easily causes error accumulation over time. Besides, warping the information at the pixel level instead of the feature level limits the performance (e.g., unable to generate different views) and is unstable since it directly demonstrates the results even with a misalignment. In contrast, fusing information at the feature level can be further refined by the convolution to obtain the final results. Based on these assumptions, we propose a co-attention feature-remapping framework, namely FashionMirror, that generates the try-on results according to the driven-pose sequence in two stages. In the first stage, we consider the source human image and the target try-on clothes to predict the removed mask and the try-on clothing mask, which replaces the pre-processed semantic segmentation and reduces the inference time. In the second stage, we first remove the clothes on the source human via the removed mask and warp the clothing features conditioning on the try-on clothing mask to fit the next frame human. Meanwhile, we predict the optical flows from the consecutive 2D poses and warp the source human to the next frame at the feature level. Then, we enhance the clothing features and source human features in every frame to generate realistic try-on results with spatio-temporal smoothness. Both qualitative and quantitative results show that FashionMirror outperforms the state-of-the-art virtual try-on approaches.

We demonstrate that it is possible to perform face-related computer vision in the wild using synthetic data alone. The community has long enjoyed the benefits of synthesizing training data with graphics, but the domain gap between real and synthetic data has remained a problem, especially for human faces. Researchers have tried to bridge this gap with data mixing, domain adaptation, and domain-adversarial training, but we show that it is possible to synthesize data with minimal domain gap, so that models trained on synthetic data generalize to real in-the-wild datasets. We describe how to combine a procedurally-generated parametric 3D face model with a comprehensive library of hand-crafted assets to render training images with unprecedented realism and diversity. We train machine learning systems for face-related tasks such as landmark localization and face parsing, showing that synthetic data can both match real data in accuracy, as well as open up new approaches where manual labeling would be impossible.

One of the most critical problems in weight-sharing neural architecture search is the evaluation of candidate models within a predefined search space. In practice, a one-shot supernet is trained to serve as an evaluator. A faithful ranking certainly leads to more accurate searching results. However, current methods are prone to making misjudgments. In this paper, we prove that their biased evaluation is due to inherent unfairness in the supernet training. In view of this, we propose two levels of constraints: expectation fairness and strict fairness. Particularly, strict fairness ensures equal optimization opportunities for all choice blocks throughout the training, which neither overestimates nor underestimates their capacity. We demonstrate that this is crucial for improving the confidence of models' ranking. Incorporating the one-shot supernet trained under the proposed fairness constraints with a multi-objective evolutionary search algorithm, we obtain various state-of-the-art models, e.g., FairNAS-A attains 77.5% top-1 validation accuracy on ImageNet.

Performing simple household tasks based on language directives is very natural to humans, yet it remains an open challenge for an AI agent. The 'interactive instruction following' task attempts to make progress towards building an agent that can jointly navigate, interact, and reason in the environment at every step. To address the multifaceted problem, we propose a model that factorizes the task into interactive perception and action policy streams with enhanced components. We empirically validate that our model outperforms prior arts by significant margins on the ALFRED benchmark in all metrics with improved generalization.

Current face image retrieval solutions are limited, since they treat different facial attributes the same and cannot incorporate user's preference for a subset of attributes in their search criteria. This paper introduces a new face image retrieval framework, where the input face query is augmented by both an adjustment vector that specifies the desired modifications to the facial attributes, and a preference vector that assigns different levels of importance to different attributes. For example, a user can ask for retrieving images similar to a query image, but with a different hair color, and no preference for absence/presence of eyeglasses in the results. To achieve this, we propose to disentangle the semantics, corresponding to various attributes, by learning a set of sparse and orthogonal basis vectors in the latent space of StyleGAN. Such basis vectors are then employed to decompose the dissimilarity between face images in terms of dissimilarity between their attributes, assign preference to the attributes, and adjust the attributes in the query. Enforcing sparsity on the basis vectors helps us to disentangle the latent space and adjust each attribute independently from other attributes, while enforcing orthogonality facilitates preference assignment and the dissimilarity decomposition. The effectiveness of our approach is illustrated by achieving state-of-the-art results for the face image retrieval task.

Recent advancements in deep neural networks have made remarkable leap-forwards in dense image prediction. However, the issue of feature alignment remains as neglected by most existing approaches for simplicity. Direct pixel addition between upsampled and local features leads to feature maps with misaligned contexts that, in turn, translate to mis-classifications in prediction, especially on object boundaries. In this paper, we propose a feature alignment module that learns transformation offsets of pixels to contextually align upsampled higher-level features; and another feature selection module to emphasize the lower-level features with rich spatial details. We then integrate these two modules in a top-down pyramidal architecture and present the Feature-aligned Pyramid Network (FaPN). Extensive experimental evaluations on four dense prediction tasks and four datasets have demonstrated the efficacy of FaPN, yielding an overall improvement of 1.2 - 2.6 points in AP / mIoU over FPN when paired with Faster / Mask R-CNN. In particular, our FaPN achieves the state-of-the-art of 56.7% mIoU on ADE20K when integrated within Mask-Former. The code is available from https://github.com/EMI-Group/FaPN.

Generalized zero-shot learning (GZSL) has achieved significant progress, with many efforts dedicated to overcoming the problems of visual-semantic domain gaps and seen-unseen bias. However, most existing methods directly use feature extraction models trained on ImageNet alone, ignoring the cross-dataset bias between ImageNet and GZSL benchmarks. Such a bias inevitably results in poor-quality visual features for GZSL tasks, which potentially limits the recognition performance on both seen and unseen classes. In this paper, we propose a simple yet effective GZSL method, termed feature refinement for generalized zero-shot learning (FREE), to tackle the above problem. FREE employs a feature refinement (FR) module that incorporates semantic-visual mapping into a unified generative model to refine the visual features of seen and unseen class samples. Furthermore, we propose a self-adaptive margin center loss (SAMC-loss) that cooperates with a semantic cycle-consistency loss to guide FR to learn class- and semantically-relevant representations, and concatenate the features in FR to extract the fully refined features. Extensive experiments on five benchmark datasets demonstrate the significant performance gain of FREE over current state-of-the-art methods and its baseline. The code is available at https://github.com/shiming-chen/FREE.

Efficient processing of high-resolution video streams is safety-critical for many robotics applications such as autonomous driving. Image downsampling is a commonly adopted technique to ensure the latency constraint is met. However, this naive approach greatly restricts an object detector's capability to identify small objects. In this paper, we propose an attentional approach that elastically magnifies certain regions while maintaining a small input canvas. The magnified regions are those that are believed to have a high probability of containing an object, whose signal can come from a dataset-wide prior or frame-level prior computed from recent object predictions. The magnification is implemented by a KDE-based mapping to transform the bounding boxes into warping parameters, which are then fed into an image sampler with anti-cropping regularization. The detector is then fed with the warped image and we apply a differentiable backward mapping to get bounding box outputs in the original space. Our regional magnification allows algorithms to make better use of high-resolution input without incurring the cost of high-resolution processing. On the autonomous driving datasets Argoverse-HD and BDD100K, we show our proposed method boosts the detection AP over standard Faster R-CNN, with and without finetuning. Additionally, building on top of the previous state-of-the-art in streaming detection, our method sets a new record for streaming AP on Argoverse-HD (from 17.8 to 23.0 on a GTX 1080 Ti GPU), suggesting that it has achieved a superior accuracy-latency tradeoff.

We propose the first learning-based approach for fast moving objects detection. Such objects are highly blurred and move over large distances within one video frame. Fast moving objects are associated with a deblurring and matting problem, also called deblatting. We show that the separation of deblatting into consecutive matting and deblurring allows achieving real-time performance, i.e. an order of magnitude speed-up, and thus enabling new classes of application. The proposed method detects fast moving objects as a truncated distance function to the trajectory by learning from synthetic data. For the sharp appearance estimation and accurate trajectory estimation, we propose a matting and fitting network that estimates the blurred appearance without background, followed by an energy minimization based deblurring. The state-of-the-art methods are outperformed in terms of recall, precision, trajectory estimation, and sharp appearance reconstruction. Compared to other methods, such as deblatting, the inference is of several orders of magnitude faster and allows applications such as real-time fast moving object detection and retrieval in large video collections.

Intelligent agents with visual sensors are allowed to actively explore their observations for better recognition performance. This task is referred to as Active Recognition (AR). Currently, most methods toward AR are implemented under a fixed-category setting, which constrains their applicability in realistic scenarios that need to incrementally learn new classes without retraining from scratch. Further, collecting massive data for novel categories is expensive. To address this demand, in this paper, we propose a unified framework towards Few-sample Lifelong Active Recognition (FLAR), which aims at performing active recognition on progressively arising novel categories that only have few training samples. Three difficulties emerge with FLAR: the lifelong recognition policy learning, the knowledge preservation of old categories, and the lack of training samples. To this end, our approach integrates prototypes, a robust representation for limited training samples, into a reinforcement learning solution, which motivates the agent to move towards views resulting in more discriminative features. Catastrophic forgetting during lifelong learning is then alleviated with knowledge distillation. Extensive experiments across two datasets, respectively for object and scene recognition, demonstrate that even without large training samples, the proposed approach could learn to actively recognize novel categories in a class-incremental behavior.

Driving requires interacting with road agents and predicting their future behaviour in order to navigate safely. We present FIERY: a probabilistic future prediction model in bird's-eye view from monocular cameras. Our model predicts future instance segmentation and motion of dynamic agents that can be transformed into non-parametric future trajectories. Our approach combines the perception, sensor fusion and prediction components of a traditional autonomous driving stack by estimating bird's-eye-view prediction directly from surround RGB monocular camera inputs. FIERY learns to model the inherent stochastic nature of the future solely from camera driving data in an end-to-end manner, without relying on HD maps, and predicts multimodal future trajectories. We show that our model outperforms previous prediction baselines on the NuScenes and Lyft datasets. The code and trained models are available at https://github.com/wayveai/fiery.

An optimal transportation map finds the most economical way to transport one probability measure to the other. It has been applied in a broad range of applications in vision, deep learning and medical images. By Brenier theory, computing the optimal transport map is equivalent to solving a Monge-Ampere equation. Due to the highly non-linear nature, the computation of optimal transportation maps in large scale is very challenging. This work proposes a simple but powerful method, the FFT-OT algorithm, to tackle this difficulty based on three key ideas. First, solving Monge-Ampere equation is converted to a fixed point problem; Second, the obliqueness property of optimal transportation maps are reformulated as Neumann boundary conditions on rectangular domains; Third, FFT is applied in each iteration to solve a Poisson equation in order to improve the efficiency. Experiments on surfaces captured from 3D scanning and reconstructed from medical imaging are conducted, and compared with other existing methods. Our experimental results show that the proposed FFT-OT algorithm is simple, general and scalable with high efficiency and accuracy.

Ternary Neural Networks (TNNs) have received much attention due to being potentially orders of magnitude faster in inference, as well as more power efficient, than full-precision counterparts. However, 2 bits are required to encode the ternary representation with only 3 quantization levels leveraged. As a result, conventional TNNs have similar memory consumption and speed compared with the standard 2-bit models, but have worse representational capability. Moreover, there is still a significant gap in accuracy between TNNs and full-precision networks, hampering their deployment to real applications. To tackle these two challenges, in this work, we first show that, under some mild constraints, computational complexity of the ternary inner product can be reduced by 2x. Second, to mitigate the performance gap, we elaborately design an implementation-dependent ternary quantization algorithm. The proposed framework is termed Fast and Accurate Ternary Neural Networks (FATNN). Experiments on image classification demonstrate that our FATNN surpasses the state-of-the-arts by a significant margin in accuracy. More importantly, speedup evaluation compared with various precisions is analyzed on several platforms, which serves as a strong benchmark for further research.

Recent methods for long-tailed instance segmentation still struggle on rare object classes with few training data. We propose a simple yet effective method, Feature Augmentation and Sampling Adaptation (FASA), that addresses the data scarcity issue by augmenting the feature space especially for rare classes. Both the Feature Augmentation (FA) and feature sampling components are adaptive to the actual training status -- FA is informed by the feature mean and variance of observed real samples from past iterations, and we sample the generated virtual features in a loss-adapted manner to avoid over-fitting. FASA does not require any elaborate loss design, and removes the need for inter-class transfer learning that often involves large cost and manually-defined head/tail class groups. We show FASA is a fast, generic method that can be easily plugged into standard or long-tailed segmentation frameworks, with consistent performance gains and little added cost. FASA is also applicable to other tasks like long-tailed classification with state-of-the-art performance.

In this paper, we propose a talking face generation method that takes an audio signal as input and a short target video clip as reference, and synthesizes a photo-realistic video of the target face with natural lip motions, head poses, and eye blinks that are in-sync with the input audio signal. We note that the synthetic face attributes include not only explicit ones such as lip motions that have high correlations with speech, but also implicit ones such as head poses and eye blinks that have only weak correlation with the input audio. To model such complicated relationships among different face attributes with input audio, we propose a FACe Implicit Attribute Learning Generative Adversarial Network (FACIAL-GAN), which integrates the phonetics-aware, context-aware, and identity-aware information to synthesize the 3D face animation with realistic motions of lips, head poses, and eye blinks. Then, our Rendering-to-Video network takes the rendered face images and the attention map of eye blinks as input to generate the photo-realistic output video frames. Experimental results and user studies show our method can generate realistic talking face videos with not only synchronized lip motions, but also natural head movements and eye blinks, with better qualities than the results of state-of-the-art methods.

Generative adversarial networks built from deep convolutional neural networks (GANs) lack the ability to exactly replicate the high-frequency components of natural images. To alleviate this issue, we introduce two novel training techniques called frequency dropping (F-Drop) and frequency matching (F-Match). The key idea of F-Drop is to filter out unnecessary high-frequency components from the input images of the discriminators. This simple modification prevents the discriminators from being confused by perturbations of the high-frequency components. In addition, F-Drop makes the GANs focus on fitting in the low-frequency domain, in which there are the dominant components of natural images. F-Match minimizes the difference between real and fake images in the frequency domain for generating more realistic images. F-Match is implemented as a regularization term in the objective functions of the generators; it penalizes the batch mean error in the frequency domain. F-Match helps the generators to fit in the high-frequency domain filtered out by F-Drop to the real image. We experimentally demonstrate that the combination of F-Drop and F-Match improves the generative performance of GANs in both the frequency and spatial domain on multiple image benchmarks.

To meet the space limitation of optical elements, free-form surfaces or high-order aspherical lenses are adopted in mobile cameras to compress volume. However, the application of free-form surfaces also introduces the problem of image quality mutation. Existing model-based deconvolution methods are inefficient in dealing with the degradation that shows a wide range of spatial variants over regions. And the deep learning techniques in low-level and physics-based vision suffer from a lack of accurate data. To address this issue, we develop a degradation framework to estimate the spatially variant point spread functions (PSFs) of mobile cameras. When input extreme-quality digital images, the proposed framework generates degraded images sharing a common domain with real-world photographs. Supplied with the synthetic image pairs, we design a Field-Of-View shared kernel prediction network (FOV-KPN) to perform spatial-adaptive reconstruction on real degraded photos. Extensive experiments demonstrate that the proposed approach achieves extreme-quality computational imaging and outperforms the state-of-the-art methods. Furthermore, we illustrate that our technique can be integrated into existing postprocessing systems, resulting in significantly improved visual quality.

This paper proposes an extreme structure from motion (SfM) algorithm for residential indoor panoramas that have little to no visual overlaps. Only a single panorama is present in a room for many cases, making the task infeasible for existing SfM algorithms. Our idea is to learn to evaluate the realism of room/door/window arrangements in the top-down semantic space. After using heuristics to enumerate possible arrangements based on door detections, we evaluate their realism scores, pick the most realistic arrangement, and return the corresponding camera poses. We evaluate the proposed approach on a dataset of 1029 panorama images with 286 houses. Our qualitative and quantitative evaluations show that an existing SfM approach completely fails for most of the houses. The proposed approach achieves the mean positional error of less than 1.0 meter for 47% of the houses and even 78% when considering the top five reconstructions. We will share the code and data in https://github.com/aminshabani/extreme-indoor-sfm.