ICCV 2021 Papers

Motion blur is one of the major challenges remaining for visual odometry methods. In low-light conditions where longer exposure times are necessary, motion blur can appear even for relatively slow camera motions. In this paper we present a novel hybrid visual odometry pipeline with direct approach that explicitly models and estimates the camera's local trajectory within exposure time. This allows us to actively compensate for any motion blur that occurs due to the camera motion. In addition, we also contribute a novel benchmarking dataset for motion blur aware visual odometry. In experiments we show that by directly modeling the image formation process we are able to improve robustness of the visual odometry, while keeping comparable accuracy as that for images without motion blur.

Active speaker detection requires a solid integration of multi-modal cues. While individual modalities can approximate a solution, accurate predictions can only be achieved by explicitly fusing the audio and visual features and modeling their temporal progression. Despite its inherent muti-modal nature, current methods still focus on modeling and fusing short-term audiovisual features for individual speakers, often at frame level. In this paper we present a novel approach to active speaker detection that directly addresses the multi-modal nature of the problem, and provides a straightforward strategy where independent visual features from potential speakers in the scene are assigned to a previously detected speech event. Our experiments show that, an small graph data structure built from local information, allows to approximate an instantaneous audio-visual assignment problem. Moreover, the temporal extension of this initial graph achieves a new state-of-the-art performance on the AVA-ActiveSpeaker dataset with a mAP of 88.8%.

Virtual 3D try-on can provide an intuitive and realistic view for online shopping and has a huge potential commercial value. However, existing 3D virtual try-on methods mainly rely on annotated 3D human shapes and garment templates, which hinders their applications in practical scenarios. 2D virtual try-on approaches provide a faster alternative to manipulate clothed humans, but lack the rich and realistic 3D representation. In this paper, we propose a novel Monocular-to-3D Virtual Try-On Network (M3D-VTON) that builds on the merits of both 2D and 3D approaches. By integrating 2D information efficiently and learning a mapping that lifts the 2D representation to 3D, we make the first attempt to reconstruct a 3D try-on mesh only taking the target clothing and a person image as inputs. The proposed M3D-VTON includes three modules: 1) The Monocular Prediction Module (MPM) that estimates an initial full-body depth map and accomplishes 2D clothes-person alignment through a novel two-stage warping procedure; 2) The Depth Refinement Module (DRM) that refines the initial body depth to produce more detailed pleat and face characteristics; 3) The Texture Fusion Module (TFM) that fuses the warped clothing with the non-target body part to refine the results. We also construct a high-quality synthesized Monocular-to-3D virtual try-on dataset, in which each person image is associated with a front and a back depth map. Extensive experiments demonstrate that the proposed M3D-VTON can manipulate and reconstruct the 3D human body wearing the given clothing with compelling details and is more efficient than other 3D approaches.

This presentation addresses the problem of reconstructing a high-resolution image from multiple lower-resolution snapshots captured from slightly different viewpoints in space and time. Key challenges for solving this super-resolution problem include (i) aligning the input pictures with sub-pixel accuracy, (ii) handling raw (noisy) images for maximal faithfulness to native camera data, and (iii) designing/learning an image prior (regularizer) well suited to the task. We address these three challenges with a hybrid algorithm building on the insight from Wronski et al. that aliasing is an ally in this setting, with parameters that can be learned end to end, while retaining the interpretability of classical approaches to inverse problems. The effectiveness of our approach is demonstrated on synthetic and real image bursts, setting a new state of the art on several benchmarks and delivering excellent qualitative results on real raw bursts captured by smartphones and prosumer cameras.

For machine learning models trained with limited labeled training data, validation stands to become the main bottleneck to reducing overall annotation costs. We propose a statistical validation algorithm that accurately estimates the F-score of binary classifiers for rare categories, where finding relevant examples to evaluate on is particularly challenging. Our key insight is that simultaneous calibration and importance sampling enables accurate estimates even in the low-sample regime (<300 samples). Critically, we also derive an accurate single-trial estimator of the variance of our method and demonstrate that this estimator is empirically accurate at low sample counts, enabling a practitioner to know how well they can trust a given low-sample estimate. When validating state-of-the-art semi-supervised models on ImageNet and iNaturalist2017, our method achieves the same estimates of model performance with up to 10x fewer labels than competing approaches. In particular, we can estimate model F1 scores with a variance of 0.005 using as few as 100 labels.

This paper focuses on the problem of low-rank tensor completion, the goal of which is to recover an underlying low-rank tensor from incomplete observations. Our method is motivated by the recently proposed t-product based on any invertible linear transforms. First, we define the new tensor average rank under the invertible real linear transforms. We then propose a new tensor completion model using a nonconvex surrogate to approximate the tensor average rank. This surrogate overcomes the discontinuity of the tensor average rank and alleviates the bias problem caused by the convex relaxation. Further, we develop an efficient algorithm to solve the proposed model and establish its convergence. Finally, experimental results on both synthetic and real data demonstrate the superiority of our method.

Adversarial training is one of the most effective defenses against adversarial attacks. Previous works suggest that overfitting is a dominant phenomenon in adversarial training leading to a large generalization gap between test and train accuracy in neural networks. In this work, we show that the observed generalization gap is closely related to the choice of the activation function. In particular, we show that using activation functions with low (exact or approximate) curvature values has a regularization effect that significantly reduces both the standard and robust generalization gaps in adversarial training. We observe this effect for both differentiable/smooth activations such as SiLU as well as non-differentiable/non-smooth activations such as LeakyReLU. In the latter case, the "approximate" curvature of the activation is low. Finally, we show that for activation functions with low curvature, the double descent phenomenon for adversarially trained models does not occur.

Gaze following, i.e., detecting the gaze target of a human subject, in 2D images has become an active topic in computer vision. However, it usually suffers from the out of frame issue due to the limited field-of-view (FoV) of 2D images. In this paper, we introduce a novel task, gaze following in 360-degree images which provide an omnidirectional FoV and can alleviate the out of frame issue. We collect the first dataset, "GazeFollow360", for this task, containing around 10,000 360-degree images with complex gaze behaviors under various scenes. Existing 2D gaze following methods suffer from performance degradation in 360-degree images since they may use the assumption that a gaze target is in the 2D gaze sight line. However, this assumption is no longer true for long-distance gaze behaviors in 360-degree images, due to the distortion brought by sphere-to-plane projection. To address this challenge, we propose a 3D sight line guided dual-pathway framework, to detect the gaze target within a local region (here) and from a distant region (there), parallelly. Specifically, the local region is obtained as a 2D cone-shaped field along the 2D projection of the sight line starting at the human subject's head position, and the distant region is obtained by searching along the sight line in 3D sphere space. Finally, the location of the gaze target is determined by fusing the estimations from both the local region and the distant region. Experimental results show that our method achieves significant improvements over previous 2D gaze following methods on our GazeFollow360 dataset.

In this paper, we present LookOut, a novel autonomy system that perceives the environment, predicts a diverse set of futures of how the scene might unroll and estimates the trajectory of the SDV by optimizing a set of contingency plans over these future realizations. In particular, we learn a diverse joint distribution over multi-agent future trajectories in a traffic scene that covers a wide range of future modes with high sample efficiency while leveraging the expressive power of generative models. Unlike previous work in diverse motion forecasting, our diversity objective explicitly rewards sampling future scenarios that require distinct reactions from the self-driving vehicle for improved safety. Our contingency planner then finds comfortable and non-conservative trajectories that ensure safe reactions to a wide range of future scenarios. Through extensive evaluations, we show that our model demonstrates significantly more diverse and sample-efficient motion forecasting in a large-scale self-driving dataset as well as safer and less conservative motion plans in long-term closed-loop simulations when compared to current state-of-the-art models.

Research in unpaired video translation has mainly focused on short-term temporal consistency by conditioning on neighboring frames. However for transfer from simulated to photorealistic sequences, available information on the underlying geometry offers potential for achieving global consistency across views. We propose a novel approach which combines unpaired image translation with neural rendering to transfer simulated to photorealistic surgical abdominal scenes. By introducing global learnable textures and a lighting-invariant view-consistency loss, our method produces consistent translations of arbitrary views and thus enables long-term consistent video synthesis. We design and test our model to generate video sequences from minimally-invasive surgical abdominal scenes. Because labeled data is often limited in this domain, photorealistic data where ground truth information from the simulated domain is preserved is especially relevant. By extending existing image-based methods to view-consistent videos, we aim to impact the applicability of simulated training and evaluation environments for surgical applications. Code and data: http://opencas.dkfz.de/video-sim2real.

Self-supervised video representation methods typically focus on the representation of temporal attributes in videos. However, the role of stationary versus non-stationary attributes is less explored: Stationary features, which remain similar throughout the video, enable the prediction of video-level action classes. Non-stationary features, which represent temporally varying attributes, are more beneficial for downstream tasks involving more fine-grained temporal understanding, such as action segmentation. We argue that a single representation to capture both types of features is sub-optimal, and propose to decompose the representation space into stationary and non-stationary features via contrastive learning from long and short views, i.e. long video sequences and their shorter sub-sequences. Stationary features are shared between the short and long views, while non-stationary features aggregate the short views to match the corresponding long view. To empirically verify our approach, we demonstrate that our stationary features work particularly well on an action recognition downstream task, while our non-stationary features perform better on action segmentation. Furthermore, we analyse the learned representations and find that stationary features capture more temporally stable, static attributes, while non-stationary features encompass more temporally varying ones.

This paper proposes a novel location-aware deep-learning-based single image reflection removal method. Our network has a reflection detection module to regress a probabilistic reflection confidence map, taking multi-scale Laplacian features as inputs. This probabilistic map tells if a region is reflection-dominated or transmission-dominated, and it is used as a cue for the network to control the feature flow when predicting the reflection and transmission layers. We design our network as a recurrent network to progressively refine reflection removal results at each iteration. The novelty is that we leverage Laplacian kernel parameters to emphasize the boundaries of strong reflections. It is beneficial to strong reflection detection and substantially improves the quality of reflection removal results. Extensive experiments verify the superior performance of the proposed method over state-of-the-art approaches. Our code and the pre-trained model can be found at https://github.com/zdlarr/Location-aware-SIRR.

As a representative of multiple kernel clustering (MKC), simple multiple kernel k-means (SimpleMKKM) is recently put forward to boosting the clustering performance by optimally fusing a group of pre-specified kernel matrices. Despite achieving significant improvement in a variety of applications, we find out that SimpleMKKM could indiscriminately force all sample pairs to be equally aligned with the same ideal similarity. As a result, it does not sufficiently take the variation of samples into consideration, leading to unsatisfying clustering performance. To address these issues, this paper proposes a novel MKC algorithm with a "local" kernel alignment, which only requires that the similarity of a sample to its k-nearest neighbours be aligned with the ideal similarity matrix. Such an alignment helps the clustering algorithm to focus on closer sample pairs that shall stay together and avoids involving unreliable similarity evaluation for farther sample pairs. After that, we theoretically show that the objective of SimpleMKKM is a special case of this local kernel alignment criterion with normalizing each base kernel matrix. Based on this observation, the proposed localized SimpleMKKM can be readily implemented by existing SimpleMKKM package. Moreover, we conduct extensive experiments on several widely used benchmark datasets to evaluate the clustering performance of localized SimpleMKKM. The experimental results have demonstrated that our algorithm consistently outperforms the state-of-the-art ones, verifying the effectiveness of the proposed local kernel alignment criterion. The code of Localized SimpleMKKM is publicly available at: https://github.com/xinwangliu/LocalizedSMKKM.

Videos with binaural audios provide an immersive viewing experience by enabling 3D sound sensation. Recent works attempt to generate binaural audio in a multimodal learning framework using large quantities of videos with accompanying binaural audio. In contrast, we attempt a more challenging problem -- synthesizing binaural audios for a video with monaural audio in a weakly supervised setting and weakly semi-supervised setting. Our key idea is that any down-stream task that can be solved only using binaural audios can be used to provide proxy supervision for binaural audio generation, thereby reducing the reliance on explicit supervision. In this work, as a proxy-task for weak supervision, we use Sound Source Localization with only audio. We design a two-stage architecture called Localize-to-Binauralize Network (L2BNet). The first stage of L2BNet is a Stereo Generation (SG) network employed to generate two-stream audio from monaural audio using visual frame information as guidance. In the second stage, an Audio Localization (AL) network is designed to use the synthesized two-stream audio to localize sound sources in visual frames. The entire network is trained end-to-end so that the AL network provides necessary supervision for the SG network. We experimentally show that our weakly-supervised framework generates two-stream audio containing binaural cues. Through user study, we further validate that our proposed approach generates binaural-quality audio using as little as 10% of explicit binaural supervision data for the SG network.

Cross-resolution image alignment is a key problem in multiscale gigapixel photography, which requires to estimate homography matrix using images with large resolution gap. Existing deep homography methods concatenate the input images or features, neglecting the explicit formulation of correspondences between them, which leads to degraded accuracy in cross-resolution challenges. In this paper, we consider the cross-resolution homography estimation as a multimodal problem, and propose a local transformer network embedded within a multiscale structure to explicitly learn correspondences between the multimodal inputs, namely, input images with different resolutions. The proposed local transformer adopts a local attention map specifically for each position in the feature. By combining the local transformer with the multiscale structure, the network is able to capture long-short range correspondences efficiently and accurately. Experiments on both the MS-COCO dataset and real-captured cross-resolution dataset show that the proposed network outperforms existing state-of-the-art feature-based and deep-learning-based homography estimation methods, and is able to accurately align images under 10x resolution gap.

For semantic segmentation, label probabilities are often uncalibrated as they are typically only the by-product of a segmentation task. Intersection over Union (IoU) and Dice score are often used as criteria for segmentation success, while metrics related to label probabilities are not often explored. However, probability calibration approaches have been studied, which match probability outputs with experimentally observed errors. These approaches mainly focus on classification tasks, but not on semantic segmentation. Thus, we propose a learning-based calibration method that focuses on multi-label semantic segmentation. Specifically, we adopt a convolutional neural network to predict local temperature values for probability calibration. One advantage of our approach is that it does not change prediction accuracy, hence allowing for calibration as a post-processing step. Experiments on the COCO, CamVid, and LPBA40 datasets demonstrate improved calibration performance for a range of different metrics. We also demonstrate the good performance of our method for multi-atlas brain segmentation from magnetic resonance images.

Computer vision tasks such as object detection and semantic/instance segmentation rely on the painstaking annotation of large training datasets. In this paper, we propose LocTex that takes advantage of the low-cost localized textual annotations (i.e., captions and synchronized mouse-over gestures) to reduce the annotation effort. We introduce a contrastive pre-training framework between images and captions and propose to supervise the cross-modal attention map with rendered mouse traces to provide coarse localization signals. Our learned visual features capture rich semantics (from free-form captions) and accurate localization (from mouse traces), which are very effective when transferred to various downstream vision tasks. Compared with ImageNet supervised pre-training, LocTex can reduce the size of the pre-training dataset by 10x or the target dataset by 2x while achieving comparable or even improved performance on COCO instance segmentation. When provided with the same amount of annotations, LocTex achieves around 4% higher accuracy than the previous state-of-the-art "vision+language" pre-training approach on the task of PASCAL VOC image classification.

Deep metric learning has been effectively used to learn distance metrics for different visual tasks like image retrieval, clustering, etc. In order to aid the training process, existing methods either use a hard mining strategy to extract the most informative samples or seek to generate hard synthetics using an additional network. Such approaches face different challenges and can lead to biased embeddings in the former case, and (i) harder optimization (ii) slower training speed (iii) higher model complexity in the latter case. In order to overcome these challenges, we propose a novel approach that looks for optimal hard negatives (LoOp) in the embedding space, taking full advantage of each tuple by calculating the minimum distance between a pair of positives and a pair of negatives. Unlike mining-based methods, our approach considers the entire space between pairs of embeddings to calculate the optimal hard negatives. Extensive experiments combining our approach and representative metric learning losses reveal a significant boost in performance on three benchmark datasets.

Given only a few available images for a novel unseen category, few-shot image generation aims to generate more data for this category. Previous works attempt to globally fuse these images by using adjustable weighted coefficients. However, there is a serious semantic misalignment between different images from a global perspective, making these works suffer from poor generation quality and diversity. To tackle this problem, we propose a novel Local-Fusion Generative Adversarial Network (LoFGAN) for few-shot image generation. Instead of using these available images as a whole, we first randomly divide them into a base image and several reference images. Next, LoFGAN matches local representations between the base and reference images based on semantic similarities and replaces the local features with the closest related local features. In this way, LoFGAN can produce more realistic and diverse images at a more fine-grained level, and simultaneously enjoy the characteristic of semantic alignment. Furthermore, a local reconstruction loss is also proposed, which can provide better training stability and generation quality. We conduct extensive experiments on three datasets, which successfully demonstrates the effectiveness of our proposed method for few-shot image generation and downstream visual applications with limited data. Code is available at https://github.com/edward3862/LoFGAN-pytorch.