WACV 2023 Papers

Current top-notch deep learning (DL) based vision models are primarily based on exploring and exploiting the inherent correlations between training data samples and their associated labels. However, a known practical challenge is their degraded performance against "noisy" data, induced by different circumstances such as spurious correlations, irrelevant contexts, domain shift, and adversarial attacks. In this work, we incorporate this binary information of "existence of noise" as treatment into image classification tasks to improve prediction accuracy by jointly estimating their treatment effects. Motivated from causal variational inference, we propose a transformer-based architecture, Treatment Learning Causal Transformer (TLT), that uses a latent generative model to estimate robust feature representations from current observational input for noise image classification. Depending on the estimated noise level (modeled as a binary treatment factor), TLT assigns the corresponding inference network trained by the designed causal loss for prediction. We also create new noisy image datasets incorporating a wide range of noise factors (e.g., object masking, style transfer, and adversarial perturbation) for performance benchmarking. The superior performance of TLT in noisy image classification is further validated by several refutation evaluation metrics. As a by-product, TLT also improves visual salience methods for perceiving noisy images.

Unsupervised video object segmentation (VOS) aims to detect the most salient object in a video sequence at the pixel level. In unsupervised VOS, most state-of-the-art methods leverage motion cues obtained from optical flow maps in addition to appearance cues to exploit the property that salient objects usually have distinctive movements compared to the background. However, as they are overly dependent on motion cues, which may be unreliable in some cases, they cannot achieve stable prediction. To reduce this motion dependency of existing two-stream VOS methods, we propose a novel motion-as-option network that optionally utilizes motion cues. Additionally, to fully exploit the property of the proposed network that motion is not always required, we introduce a collaborative network learning strategy. On all the public benchmark datasets, our proposed network affords state-of-the-art performance with real-time inference speed.

There is evidence that transformers offer state-of-the-art recognition performance on tasks involving overhead imagery (e.g., satellite imagery). However, it is difficult to make unbiased empirical comparisons between competing deep learning models, making it unclear whether, and to what extent, transformer-based models are beneficial. In this paper we systematically compare the impact of adding transformer structures into state-of-the-art segmentation models for overhead imagery. Each model is given a similar budget of free parameters, and their hyperparameters are optimized using Bayesian Optimization with a fixed quantity of data and computation time. We conduct our experiments with a large and diverse dataset comprising two large public benchmarks: Inria and DeepGlobe. We perform additional ablation studies to explore the impact of specific transformer-based modeling choices. Our results suggest that transformers provide consistent, but modest, performance improvements. We only observe this advantage however in hybrid models that combine convolutional and transformer-based structures, while fully transformer-based models achieve relatively poor performance.

Visual geo-localization remains a challenging task due to variations in the appearance and perspective among captured images. This paper introduces an efficient TransVLAD module, which aggregates attention-based feature maps into a discriminative and compact global descriptor. Unlike existing methods that generate feature maps using only convolutional neural networks (CNNs), we propose a sparse transformer to encode global dependencies and compute attention-based feature maps, which effectively reduces visual ambiguities that occurs in large-scale geo-localization problems. A positional embedding mechanism is used to learn the corresponding geometric configurations between query and gallery images. A grouped VLAD layer is also introduced to reduce the number of parameters, and thus construct an efficient module. Finally, rather than only learning from the global descriptors on entire images, we propose a self-supervised learning method to further encode more information from multi-scale patches between the query and positive gallery images. Extensive experiments on three challenging large-scale datasets indicate that our model outperforms state-of-the-art models, and has lower computational complexity.

3D object detection using point clouds has attracted increasing attention due to its wide applications in autonomous driving and robotics. However, most existing studies focus on single point cloud frames without harnessing the temporal information in point cloud sequences. In this paper, we design TransPillars, a novel transformer-based feature aggregation technique that exploits temporal features of consecutive point cloud frames for multi-frame 3D object detection. TransPillars aggregates spatial-temporal point cloud features from two perspectives. First, it fuses voxel-level features directly from multi-frame feature maps instead of pooled instance features to preserve instance details with contextual information that are essential to accurate object localization. Second, it introduces a hierarchical coarse-to-fine strategy to fuse multi-scale features progressively to effectively capture the motion of moving objects and guide the aggregation of fine features. Besides, a variant of deformable transformer is introduced to improve the effectiveness of cross-frame feature matching. Extensive experiments show that our proposed TransPillars achieves state-of-art performance as compared to existing multi-frame detection approaches.

Tracking multiple objects in videos relies on modeling the spatial-temporal interactions of the objects. In this paper, we propose TransMOT, which leverages powerful graph transformers to efficiently model the spatial and temporal interactions among the objects. TransMOT is capable of effectively modeling the interactions of a large number of objects by arranging the trajectories of the tracked targets and detection candidates as a set of sparse weighted graphs, and constructing a spatial graph transformer encoder layer, a temporal transformer encoder layer, and a spatial graph transformer decoder layer based on the graphs. Through end-to-end learning, TransMOT can exploit the spatial-temporal clues to directly estimate association from a large number of loosely filtered detection predictions for robust MOT in complex scenes. The proposed method is evaluated on multiple benchmark datasets, including MOT15, MOT16, MOT17, and MOT20, and it achieves state-of-the-art performance on all the datasets.

Humans have an innate ability to sense their surroundings, as they can extract the spatial representation from the egocentric perception and form an allocentric semantic map via spatial transformation and memory updating. However, endowing mobile agents with such a spatial sensing ability is still a challenge, due to two difficulties: (1) the previous convolutional models are limited by the local receptive field, thus, struggling to capture holistic long-range dependencies during observation; (2) the excessive computational budgets required for success, often lead to a separation of the mapping pipeline into stages, resulting the entire mapping process inefficient. To address these issues, we propose an end-to-end one-stage Transformer-based framework for Mapping, termed Trans4Map. Our egocentric-to-allocentric mapping process includes three steps: (1) the efficient transformer extracts the contextual features from a batch of egocentric images; (2) the proposed Bidirectional Allocentric Memory (BAM) module projects egocentric features into the allocentric memory; (3) the map decoder parses the accumulated memory and predicts the top-down semantic segmentation map. In contrast, Trans4Map achieves state-of-the-art results, reducing 67.2% parameters, yet gaining a +3.25% mIoU and a +4.09% mBF1 improvements on the Matterport3D dataset.

Machine learning-based algorithms using fully convolutional networks (FCNs) have been a promising option for medical image segmentation. However, such deep networks silently fail if input samples are drawn far from the training data distribution, thus causing critical problems in automatic data processing pipelines. To overcome such out-of-distribution (OoD) problems, we propose a novel OoD score formulation and its regularization strategy by applying an auxiliary add-on classifier to an intermediate layer of an FCN, where the auxiliary module is helfpul for analyzing the encoder output features by taking their class information into account. Our regularization strategy train the module along with the FCN via the principle of outlier exposure so that our model can be trained to distinguish OoD samples from normal ones without modifying the original network architecture. Our extensive experiment results demonstrate that the proposed approach can successfully conduct effective OoD detection without loss of segmentation performance. In addition, our module can provide reasonable explanation maps along with OoD scores, which can enable users to analyze the reliability of predictions.

Plant roots are difficult to monitor and study since they are hidden belowground. Minirhizotrons offer an in-situ monitoring solution but their widespread adoption is still limited by the capabilities of automatic analysis methods. These capabilities so far consist only of estimating a single number (total root length) per image.\nWe propose a method for a more fine-grained analysis which estimates the root turnover, i.e. the amount of root growth and decay between two minirhizotron images. It consists of a neural network that computes which roots are visible in both images and is trained in an unsupervised manner without additional annotations.\nOur code is available as a part of an analysis tool with a user interface ready to be used by ecologists.

The design of pedestrian-friendly infrastructures plays a crucial role in creating sustainable transportation in urban environments. Analyzing pedestrian behaviour in response to existing infrastructure is pivotal to planning, maintaining, and creating more pedestrian-friendly facilities. Many approaches have been proposed to extract such behaviour by applying deep learning models to video data. Video data, however, includes an broad spectrum of privacy-sensitive information about individuals, such as their location at a given time or who they are with. Most of the existing models use privacy-invasive methodologies to track, detect, and analyse individual or group pedestrian behaviour patterns. As a step towards privacy-preserving pedestrian analysis, this paper introduces a framework to anonymize all pedestrians before analyzing their behaviors. The proposed framework leverages recent developments in 3D wireframe reconstruction and digital in-painting to represent pedestrians with quantitative wireframes by removing their images while preserving pose, shape, and background scene context. To evaluate the proposed framework, a generic metric is introduced for each of privacy and utility. Experimental evaluation on widely-used datasets shows that the proposed framework outperforms traditional and state-of-the-art image filtering approaches by generating best privacy utility trade-off.

Existing object detection models assume both the training and test data are sampled from the same source domain. This assumption does not hold true when these detectors are deployed in real-world applications, where they encounter new visual domains. Unsupervised Domain Adaptation (UDA) methods are generally employed to mitigate the adverse effects caused by domain shift. Existing UDA methods operate in an offline manner where the model is first adapted toward the target domain and then deployed in real-world applications. However, this offline adaptation strategy is not suitable for real-world applications as the model frequently encounters new domain shifts. Hence, it is critical to develop a feasible UDA method that generalizes to the new domain shifts encountered during deployment time in a continuous online manner. To this end, we propose a novel unified adaptation framework that adapts and improves generalization on the target domain in both offline and online settings. Specifically, we introduce MemXformer - a cross-attention transformer-based memory module where items in the memory take advantage of domain shifts and record prototypical patterns of the target distribution. Further, MemXformer produces strong positive and negative pairs to guide a novel contrastive loss, which enhances target-specific representation learning. Experiments on diverse detection benchmarks show that the proposed strategy producs state-of-the-art performance in both offline and online settings. To the best of our knowledge, this is the first work to address online and offline adaptation settings for object detection. Source code will be released after review.

Many people with some form of hearing loss consider lipreading as their primary mode of day-to-day communication. However, finding resources to learn or improve one's lipreading skills can be challenging. This is further exacerbated in the COVID19 pandemic due to restrictions on direct interactions with peers and speech therapists. Today, online MOOCs platforms like Coursera and Udemy have become the most effective form of training for many types of skill development. However, online lipreading resources are scarce as creating such resources is an extensive process needing months of manual effort to record hired actors. Because of the manual pipeline, such platforms are also limited in vocabulary, supported languages, accents, and speakers and have a high usage cost. In this work, we investigate the possibility of replacing real human talking videos with synthetically generated videos. Synthetic data can easily incorporate larger vocabularies, variations in accent, and even local languages and many speakers. We propose an end-to-end automated pipeline to develop such a platform using state-of-the-art talking head video generator networks, text-to-speech models, and computer vision techniques. We then perform an extensive human evaluation using carefully thought out lipreading exercises to validate the quality of our designed platform against the existing lipreading platforms. Our studies concretely point toward the potential of our approach in developing a large-scale lipreading MOOC platform that can impact millions of people with hearing loss.

Most video anomaly detection approaches are based on data-intensive end-to-end trained neural networks, which extract spatiotemporal features from videos. The extracted feature representations in such approaches are not interpretable, which prevents the automatic identification of anomaly cause. To this end, we propose a novel framework which can explain the detected anomalous event in a surveillance video. In addition to monitoring objects independently, we also monitor the interactions between them to detect anomalous events and explain their root causes. Specifically, we demonstrate that the scene graphs obtained by monitoring the object interactions provide an interpretation for the context of the anomaly while performing competitively with respect to the recent state-of-the-art approaches. Moreover, the proposed interpretable method enables cross-domain adaptability (i.e., transfer learning in another surveillance scene), which is not feasible for most existing end-to-end methods due to the lack of sufficient labeled training data for every surveillance scene. The quick and reliable detection performance of the proposed method is evaluated both theoretically (through an asymptotic optimality proof) and empirically on the popular benchmark datasets.

Talking-face video generation works have achieved state-of-the-art results in synthesizing videos with lip synchronization. However, most of the previous works deal with low-resolution talking-face videos (up to 256x256 pixels), thus, generating extremely high-resolution videos still remains a challenge. We take a giant leap in this work and propose a novel method to synthesize talking-face videos at resolutions as high as 4K! Our task presents several key challenges: (i) Scaling the existing methods to such high resolutions is resource-constrained, both in terms of compute and the availability of very high-resolution datasets, (ii) The synthesized videos need to be spatially and temporally coherent. The sheer number of pixels that the model needs to generate while maintaining the temporal consistency at the video level makes this task non-trivial and has never been attempted before in literature. To address these issues, we propose to train the lip-sync generator in a compact Vector Quantized (VQ) space for the first time. Our core idea to encode the faces in a compact 16x16 representation allows us to model high-resolution videos. In our framework, we learn the lip movements in the quantized space on the newly collected 4K Talking Faces (4KTF) dataset. Our approach is speaker agnostic and can handle various languages and voices. We benchmark our technique against several competitive works and show that we can achieve a remarkable 64-times more pixels than the current state-of-the-art! Our supplementary demo video depicts additional qualitative results, comparisons, and several real-world applications, like professional movie editing enabled by our model.

For specialized and dense downstream tasks such as object detection, labeling data requires expertise and can be very expensive, making few-shot and semi-supervised models much more attractive alternatives. While in the few-shot setup we observe that transformer-based object detectors perform better than convolution-based two-stage models for a similar amount of parameters, they are not as effective when used with recent approaches in the semi-supervised setting. In this paper, we propose a semi-supervised method tailored for the current state-of-the-art object detector Deformable DETR in the few-annotation learning setup using a student-teacher architecture, which avoids relying on a sensitive post-processing of the pseudo-labels generated by the teacher model. We evaluate our method on the semi-supervised object detection benchmarks COCO and Pascal VOC, and it outperforms previous methods, especially when annotations are scarce. We believe that our contributions open new possibilities to adapt similar object detection methods in this setup as well.

Methods for Optical Flow (OF) estimation based on Deep Learning have considerably improved traditional approaches in challenging and realistic conditions. However, data-driven approaches can inherently be biased, leading to unexpected under-performance in real application scenarios. In this paper, we first observe that the OF estimation accuracy varies with motion direction, and name this phenomenon 'OF sign imbalance'. The sign imbalance cannot be assessed by means of the endpoint-error (EPE), the typical training and evaluation metric for Deep Optical Flow estimators. This paper tackles this issue by proposing a new metric to assess the sign imbalance, which is compared to the endpoint-error. We provide an extensive evaluation of the sign imbalance for the state-of-the-art optical flow estimators. Based on the evaluation, we propose two strategies to mitigate the phenomenon, i) by constraining the model estimations during inference, and, ii) by constraining the loss function during training. Testing and training code is available at: www.github.com/stsavian/equivariant_of_estimation.

Deep reinforcement learning has shown promising results on an abundance of robotic tasks in simulation, including visual navigation and manipulation. Prior work generally aims to build embodied agents that solve their assigned tasks as quickly as possible, while largely ignoring the problems caused by collision with objects during interaction. This lack of prioritization is understandable: there is no inherent cost in breaking virtual objects. As a result, "well-trained" agents frequently collide with objects before achieving their primary goals, a behavior that would be catastrophic in the real world. In this paper, we study the problem of training agents to complete the task of visual mobile manipulation in the ManipulaTHOR environment while avoiding unnecessary collision (disturbance) with objects. We formulate disturbance avoidance as a penalty term in the reward function, but find that directly training with such penalized rewards often results in agents being unable to escape poor local optima. Instead, we propose a two-stage training curriculum where an agent is first allowed to freely explore and build basic competencies without penalization, after which a disturbance penalty is introduced to refine the agent's behavior. Results on testing scenes show that our curriculum not only avoids these poor local optima, but also leads to 10% absolute gains in success rate without disturbance, compared to our state-of-the-art baselines. Moreover, our curriculum is significantly more performant than a safe RL algorithm that casts collision avoidance as a constraint. Finally, we propose a novel disturbance-prediction auxiliary task that accelerates learning.

Recent object detection models have proved valuable for many robotics and manufacturing tasks, but they require large amounts of annotated data for each new class of objects they are trained for. Few-shot object detection (FSOD) aims to address this problem by learning novel classes given only a few samples of annotated data. While competitive results have been achieved using two-stage FSOD detectors, typically faster one-stage FSODs underperform in comparison. We make the discovery that the large gap in performance between two-stage and one-stage FSODs is mainly due to their weak discriminability, which is explained away by a small post-fusion receptive field and a small number of foreground samples in the loss function. We propose a new one-stage FSOD framework to address these limitations - Few-shot RetinaNet (FSRN). Specifically, we propose: (1) a multi-way support training strategy to augment the number of foreground samples for dense meta-detectors during training, (2) an early multi-level feature fusion providing a wide receptive field that covers the whole anchor area, (3) two augmentation techniques on query and source images to enhance transferability. Extensive experiments demonstrate that the proposed approach addresses the limitations of previous methods and boosts both discriminability and transferability. FSRN is two times faster than twostage FSODs while remaining competitive in accuracy, and it triples the state-of-the-art of one-stage meta-detectors on the competitive 10-shot MS-COCO benchmark. On the PASCAL VOC benchmark, the proposed approach consistently outperforms one-stage meta-detectors and many two-stage FSODs.

In this work, we propose a novel and scalable solution to address the challenges of developing efficient dense predictions on edge platforms. Our first key insight is that MultiTask Learning (MTL) and hardware-aware Neural Architecture Search (NAS) can work in synergy to greatly benefit on-device Dense Predictions (DP). Empirical results reveal that the joint learning of the two paradigms is surprisingly effective at improving DP accuracy, achieving superior performance over both the transfer learning of single-task NAS and prior state-of-the-art approaches in MTL, all with just 1/10th of the computation. To the best of our knowledge, our framework, named EDNAS, is the first to successfully leverage the synergistic relationship of NAS and MTL for DP. Our second key insight is that the standard depth training for multi-task DP can cause significant instability and noise to MTL evaluation. Instead, we propose JAReD, an improved, easy-to-adopt Joint Absolute-Relative Depth loss, that reduces up to 88% of the undesired noise while simultaneously boosting accuracy. We conduct extensive evaluations on standard datasets, benchmark against strong baselines and state-of-the-art approaches, as well as provide an analysis of the discovered optimal architectures.