WACV 2022 Papers

Unsupervised domain adaptation is often addressed by learning a joint representation of labeled samples from a source domain and unlabeled samples from a target domain. Unfortunately, hard sharing of representation may hurt adaptation because of negative transfer, where features that are useful for source domains are learned even if they hurt inference on the target domain. Here, we propose an alternative, soft sharing scheme. We train separate but weakly-coupled models for the source and the target data, while encouraging their predictions to agree. Training the two coupled models jointly effectively exploits the distribution over unlabeled target data and achieves high accuracy on the target. Specifically, we show analytically and empirically that the decision boundaries of the target model converge to low-density "valleys" of the target distribution. We evaluate our approach on four multi-source domain adaptation (MSDA) benchmarks, digits, amazon text reviews, Office-Caltech, and images (DomainNet). We find that it consistently outperforms current MSDA SoTA, sometimes by a very large margin.

In this paper, we investigate visual-based camera relocalization with neural networks for robotics and autonomous vehicles applications. Our solution is a CNN-based algorithm which predicts camera pose (3D translation and 3D rotation) directly from a single image. It also provides an uncertainty estimate of the pose. Pose and uncertainty are learned together with a single loss function and are fused at test time with an EKF. Furthermore, we propose a new fully convolutional architecture, named CoordiNet, designed to embed some of the scene geometry. Our framework outperforms comparable methods on the largest available benchmark, the Oxford RobotCar dataset, with an average error of 8 meters where previous best was 19 meters. We have also investigated the performance of our method on large scenes for real time (18 fps) vehicle localization. In this setup, structure-based methods require a large database, and we show that our proposal is a reliable alternative, achieving 29cm median error in a 1.9km loop in a busy urban area.

The state-of-the-art StyleGAN2 network supports powerful methods to create and edit art, including generating random images, finding images "like" some query, and modifying content or style. Further, recent advancements enable training with small datasets. We apply these methods to synthesize card art, by training on a novel Yu-Gi-Oh dataset. While noise inputs to StyleGAN2 are essential for good synthesis, we find that coarse-scale noise interferes with latent variables on this dataset because both control long-scale image effects. We observe over-aggressive variation in art with changes in noise and weak content control via latent variable edits. Here, we demonstrate that training a modified StyleGAN2, where coarse-scale noise is suppressed, removes these unwanted effects. We obtain a superior FID; changes in noise result in local exploration of style; and identity control is markedly improved. These results and analysis lead towards a GAN-assisted art synthesis tool for digital artists of all skill levels, which can be used in film, games, or any creative industry for artistic ideation.

The success of learning with noisy labels (LNL) methods relies heavily on the success of a warm-up stage where standard supervised training is performed using the full (noisy) training set. In this paper, we identify a "warm-up obstacle": the inability of standard warm-up stages to train high quality feature extractors and avert memorization of noisy labels. We propose "Contrast to Divide" (C2D), a simple framework that solves this problem by pre-training the feature extractor in a self-supervised fashion. Using self-supervised pre-training boosts the performance of existing LNL approaches by drastically reducing the warm-up stage's susceptibility to noise level, shortening its duration, and improving extracted feature quality. C2D works out of the box with existing methods and demonstrates markedly improved performance, especially in the high noise regime, where we get a boost of more than 27% for CIFAR-100 with 90% noise over the previous state of the art. In real-life noise settings, C2D trained on mini-WebVision outperforms previous works both in WebVision and ImageNet validation sets by 3% top-1 accuracy. We perform an in-depth analysis of the framework, including investigating the performance of different pre-training approaches and estimating the effective upper bound of the LNL performance with semi-supervised learning. Code for reproducing our experiments is available at https://github.com/ContrastToDivide/C2D

This paper investigates the problem of Temporal Action Proposal (TAP) generation, which aims to provide a set of high-quality video segments that potentially contain actions events locating in long untrimmed videos. Based on the goal to distill available contextual information, we introduce a Contextual Proposal Network (CPN) composing of two context-aware mechanisms. The first mechanism, i.e., feature enhancing, integrates the inception-like module with long-range attention to capture the multi-scale temporal contexts for yielding a robust video segment representation. The second mechanism, i.e., boundary scoring, employs the bi-directional recurrent neural networks (RNN) to capture bi-directional temporal contexts that explicitly model actionness, background, and confidence of proposals. While generating and scoring proposals, such bi-directional temporal contexts are helpful to retrieve high-quality proposals of low false positives for covering the video action instances. We conduct experiments on two challenging datasets of ActivityNet-1.3 and THUMOS-14 to demonstrate the effectiveness of the proposed Contextual Proposal Network (CPN). In particular, our method respectively surpasses state-of-the-art TAP methods by 1.54% AUC on ActivityNet-1.3 test split and by 0.61% AR@200 on THUMOS-14 dataset.

Model-agnostic meta-learning (MAML) is a well-known optimization-based meta-learning algorithm that works well in various computer vision tasks, e.g., few-shot classification. MAML is to learn an initialization so that a model can adapt to a new task in a few steps. However, since the gradient norm of a classifier (head) is much bigger than those of backbone layers, the model focuses on learning the decision boundary of the classifier with similar representations. Furthermore, gradient norms of high-level layers are small than those of the other layers. So, the backbone of MAML usually learns task-generic features, which results in deteriorated adaptation performance in the inner-loop. To resolve or mitigate this problem, we propose contextual gradient scaling (CxGrad), which scales gradient norms of the backbone to facilitate learning task-specific knowledge in the inner-loop. Since the scaling factors are generated from task-conditioned parameters, gradient norms of the backbone can be scaled in a task-wise fashion. Experimental results show that CxGrad effectively encourages the backbone to learn task-specific knowledge in the inner-loop and improves the performance of MAML up to a significant margin in both same- and cross-domain few-shot classification.

Multi-object tracking is essential in biomedical image analysis. Most multi-object tracking methods follow a tracking-by-detection approach that involves using object detectors and learning the appearance feature models of the detected regions for association. Although these methods can learn the appearance similarity features to identify the same objects among frames, they have difficulties identifying the same cells because cells have a similar appearance and their shapes change as they migrate. In addition, cells often partially overlap for several frames. In this case, even an expert biologist would require knowledge of the spatial-temporal context in order to identify individual cells. To tackle such difficult situations, we propose a cell-tracking method that can effectively use the spatial-temporal context in multiple frames by using long-term motion estimation and an object-level warping loss. We conducted experiments showing that the proposed method outperformed state-of-the-art methods under various conditions on real biological images.

Compressed sensing (CS) is extensively used to reduce magnetic resonance imaging (MRI) acquisition time. State-of-the-art deep learning-based methods have proven effective in obtaining fast, high-quality reconstruction of CS-MR images. However, they treat the inherently complex-valued MRI data as real-valued entities by extracting the magnitude content or concatenating the complex-valued data as two real-valued channels for processing. In both cases, the phase content is discarded. To address the fundamental problem of real-valued deep networks, i.e. their inability to process complex-valued data, we propose a complex-valued generative adversarial network (Co-VeGAN) framework, which is the first-of-its-kind generative model exploring the use of complex-valued weights and operations. Further, since real-valued activation functions do not generalize well to the complex-valued space, we propose a novel complex-valued activation function that is sensitive to the input phase and has a learnable profile. Extensive evaluation of the proposed approach on different datasets demonstrates that it significantly outperforms the existing CS-MRI reconstruction techniques.

Although various face-related tasks have significantly advanced in recent years, occlusion and extreme pose still impede the achievement of higher performance. Existing face rotation or de-occlusion methods only have emphasized the aspect of each problem. In addition, the lack of high-quality paired data remains an obstacle for both methods. In this work, we present a self-supervision strategy called Swap-R&R to overcome the lack of ground-truth in a fully unsupervised manner for joint face rotation and de-occlusion. To generate an input pair for self-supervision, we transfer the occlusion from a face in an image to an estimated 3D face and create a damaged face image, as if rotated from a different pose by rotating twice with the roughly de-occluded face. Furthermore, we propose Complete Face Recovery GAN (CFR-GAN) to restore the collapsed textures and disappeared occlusion areas by leveraging the structural and textural differences between two rendered images. Unlike previous works, which have selected occlusion-free images to obtain ground-truths, our approach does not require human intervention and paired data. We show that our proposed method can generate a de-occluded frontal face image from an occluded profile face image. Moreover, extensive experiments demonstrate that our approach can boost the performance of facial recognition and facial expression recognition. The code is publicly available.

Tracking by detection paradigm is one of the most popular object tracking methods. However, it is very dependent on the performance of the detector. When the detector has a behavior of missing detection, the tracking result will be directly affected. In this paper, we analyze the phenomenon of the lost tracking object in real-time tracking model on MOT2020 dataset. Based on simple and traditional methods, we propose a compensation tracker to further alleviate the lost tracking problem caused by missing detection. It consists of a motion compensation module and an object selection module. The proposed method not only can re-track missing tracking objects from lost objects, but also does not require additional networks so as to maintain speed-accuracy trade-off of the real-time model. Our method only needs to be embedded into the tracker to work without re-training the network. Experiments show that the compensation tracker can efficaciously improve the performance of the model and reduce identity switches. With limited costs, the compensation tracker successfully enhances the baseline tracking performance by a large margin and reaches 66% of MOTA and 67% of IDF1 on MOT2020 dataset.

The goal of video captioning is to generate captions for a video by understanding visual and temporal cues. A general video captioning model consists of an Encoder-Decoder framework where Encoder generally captures the visual and temporal information while the decoder generates captions. Recent works have incorporated object-level information into the Encoder by a pretrained off-the-shelf object detector, significantly improving performance. However, using an object detector comes with the following downsides: 1) object detectors may not exhaustively capture all the object categories. 2) In a realistic setting, the performance may be influenced by the domain gap between the object detector and the visual-captioning dataset. To remedy this, we argue that using an external object detector could be eliminated if the model is equipped with the capability of automatically finding salient regions. To achieve this, we propose a novel architecture that learns to attend to salient regions such as objects, persons automatically using a co-segmentation inspired attention module. Then, we utilize a novel salient region interaction module to promote information propagation between salient regions of adjacent frames. Further, we incorporate this salient region-level information into the model using knowledge distillation. We evaluate our model on two benchmark datasets MSR-VTT and MSVD, and show that our model achieves competitive performance without using any object detector.

In this paper, a fine-to-finer segmentation task is investigated driven by region and contour features collaboratively on Glomerular Electron-Dense Deposits (GEDD) in view of the complementary nature of these two types of features. To this end, a novel network (Co-Net) is presented to dynamically use fine saliency segmentation to guide finer segmentation on boundaries. The whole architecture contains double mutually boosted decoders sharing one common encoder. Specifically, a new structure named Global-guided Interaction Module (GIM) is designed to effectively control the information flow and reduce redundancy in the cross-level feature fusion process. At the same time, the global features are used in it to make the features of each layer gain access to richer context, and a fine segmentation map is obtained initially; Discontinuous Boundary Supervision (DBS) strategy is applied to pay more attention to discontinuity positions and modifying segmentation errors on boundaries. At last, Selective Kernel (SK) is used for dynamical aggregation of the region and contour features to obtain a finer segmentation. Our proposed approach is evaluated on an independent GEDD dataset labeled by pathologists and also on open polyp datasets to test the generalization. Ablation studies show the effectiveness of different modules. On all datasets, our proposal achieves high segmentation accuracy and surpasses previous methods.

Conventional Unsupervised Domain Adaptation (UDA) methods presume source and target domain data to be simultaneously available during training. Such an assumption may not hold in practice, as source data is often inaccessible (e.g., due to privacy reasons). On the contrary, a pre-trained source model is usually available, which performs poorly on target due to the well-known domain shift problem. This translates into a significant amount of misclassifications, which can be interpreted as structured noise affecting the inferred target pseudo-labels. In this work, we cast UDA as a pseudo-label refinery problem in the challenging source-free scenario. We propose Negative Ensemble Learning (NEL) technique, a unified method for adaptive noise filtering and progressive pseudo-label refinement. NEL is devised to tackle noisy pseudo-labels by enhancing diversity in ensemble members with different stochastic (i) input augmentation and (ii) feedback. The latter is achieved by leveraging the novel concept of Disjoint Residual Labels, which allow propagating diverse information to the different members. Eventually, a single model is trained with the refined pseudo-labels, which leads to a robust performance on the target domain. Extensive experiments show that the proposed method achieves state-of-the-art performance on major UDA benchmarks, such as Digit5, PACS, Visda-C, and DomainNet, without using source data samples at all.

Active learning aims to reduce the labeling effort that is required to train algorithms by learning an acquisition function selecting the most relevant data for which a label should be requested from a large unlabeled data pool. Active learning is generally studied on balanced datasets where an equal amount of images per class is available. However, real-world datasets suffer from severe imbalanced classes, the so called long-tail distribution. We argue that this further complicates the active learning process, since the imbalanced data pool can result in suboptimal classifiers. To address this problem in the context of active learning, we proposed a general optimization framework that explicitly takes class-balancing into account. Results on three datasets showed that the method is general (it can be combined with most existing active learning algorithms) and can be effectively applied to boost the performance of both informative and representative-based active learning methods. In addition, we showed that also on balanced datasets our method generally results in a performance gain.

We introduce CharacterGAN, a generative model that can be trained on only a few samples (8 - 15) of a given character. Our model generates novel poses based on keypoint locations, which can be modified in real time while providing interactive feedback, allowing for intuitive reposing and animation. Since we only have very limited training samples, one of the key challenges lies in how to address (dis)occlusions, e.g. when a hand moves behind or in front of a body. To address this, we introduce a novel layering approach which explicitly splits the input keypoints into different layers which are processed independently. These layers represent different parts of the character and provide a strong implicit bias that helps to obtain realistic results even with strong (dis)occlusions. To combine the features of individual layers we use an adaptive scaling approach conditioned on all keypoints. Finally, we introduce a mask connectivity constraint to reduce distortion artifacts that occur with extreme out-of-distribution poses at test time. We show that our approach outperforms recent baselines and creates realistic animations for diverse characters. We also show that our model can handle discrete state changes, for example a profile facing left or right, that the different layers do indeed learn features specific for the respective keypoints in those layers, and that our model scales to larger datasets when more data is available. Code is available at https://github.com/tohinz/CharacterGAN.

Channel pruning has become an effective yet still challenging approach to achieve compact neural networks. It aims to prune the optimal set of filters whose removal results in minimal performance degradation of the slimmed network. Due to the prohibitively vast search space of filter combinations, existing approaches usually use various criteria to estimate the filter importance while sacrificing some precision. Here we present a new approach to optimizing the filter selection in channel pruning with lookahead search guided reinforcement learning (RL). A neural network that takes as input filter-related features is trained with RL to prune the optimal sequence of filters and maximize the performance of the remaining network. In addition, we employ Monte Carlo tree search (MCTS) to provide a lookahead search for filter selection, which increases the sample efficiency for the RL training. Experiments on MNIST, CIFAR-10, and ILSVRC-2012 validate the effectiveness of our approach compared to both traditional and automated existing channel pruning approaches.

We focus on Multimodal Machine Reading Comprehension (M3C) where a model is expected to answer questions based on given passage (or context), and the context and the questions can be in different modalities. Previous works such as RecipeQA have proposed datasets and cloze-style tasks for evaluation. However, we identify three critical biases stemming from the question-answer generation process and memorization capabilities of large deep models. These biases makes it easier for a model to overfit by relying on spurious correlations or naive data patterns. We propose a systematic framework to address these biases through three Control-Knobs that enable us to generate a test bed of datasets of progressive difficulty levels. We believe that our benchmark (referred to as Meta- RecipeQA) will provide, for the first time, a fine grained estimate of a model's generalization capabilities. We also propose a generalM3C model that is used to realize several prior SOTA models and motivate a novel hierarchical transformer based reasoning network (HTRN). We perform a detailed evaluation of these models with different language and visual features on our benchmark. We observe a consistent improvement with HTRN over SOTA ( 18% in Visual Cloze task and 13% in average over all the tasks). We also observe a drop in performance across all the models when testing on RecipeQA and proposed Meta-RecipeQA (e.g. 83.6% versus 67.1% for HTRN), which shows that the proposed dataset is relatively less biased. We conclude by highlighting the impact of the control knobs with some quantitative results.

In this paper, we introduce a novel road marking benchmark dataset for road marking detection, addressing the limitations in the existing publicly available datasets such as lack of challenging scenarios, prominence given to lane markings, unavailability of an evaluation script, lack of annotation formats and lower resolutions. Our dataset consists of 2887 total images with 4706 road marking instances belonging to 11 classes. The images have a high resolution of 1920 x 1080 and capture a wide range of traffic, lighting and weather conditions. We provide road marking annotations in polygons, bounding boxes and pixel-level segmentation masks to facilitate a diverse range of road marking detection algorithms. The evaluation metrics and the evaluation script we provide, will further promote direct comparison of novel approaches for road marking detection with existing methods. Furthermore, we evaluate the effectiveness of using both instance segmentation and object detection based approaches for the road marking detection task. Speed and accuracy scores for two instance segmentation models and two object detector models are provided as a performance baseline for our benchmark dataset. The dataset and the evaluation script is publicly available.

Although few-shot meta learning has been extensively studied in machine learning community, the fast adaptation towards new tasks remains a challenge in the few-shot learning scenario. The neuroscience research reveals that the capability of evolving neural network formulation is essential for task adaptation, which has been broadly studied in recent meta-learning researches. In this paper, we present a novel forward-backward meta-learning framework (FBM) to facilitate the model generalization in few-shot learning from a new perspective, i.e., neuron calibration. In particular, FBM models the neurons in deep neural network-based model as calibrated units under a general formulation, where neuron calibration could empower fast adaptation capability to the neural network-based models through influencing both their forward inference path and backward propagation path. The proposed calibration scheme is lightweight and applicable to various feed-forward neural network architectures. Extensive empirical experiments on the challenging few-shot learning benchmarks validate that our approach training with neuron calibration achieves a promising performance, which demonstrates that neuron calibration plays a vital role in improving the few-shot learning performance.