WACV 2023 Papers

Recent neural architecture search (NAS) works proposed training-free metrics to rank networks which largely reduced the search cost in NAS. In this paper, we revisit these training-free metrics and find that: (1) the number of parameters (#Param), which is the most straightforward training-free metric, is overlooked in previous works but is surprisingly effective, (2) recent training-free metrics largely rely on the #Param information to rank networks. Our experiments show that the performance of recent training-free metrics drops dramatically when the #Param information is not available. Motivated by these observations, we argue that metrics less correlated with the #Param are desired to provide additional information for NAS. We propose a light-weight training-based metric which has a weak correlation with the #Param while achieving better performance than training-free metrics at a lower search cost. Specifically, on DARTS search space, our method completes searching directly on ImageNet in only 2.6 GPU hours and achieves a top-1/top-5 error rate of 24.1%/7.1%, which is competitive among state-of-the-art NAS methods.

Pose estimation of the human body and hands is a fundamental problem in computer vision, and learning-based solutions require a large amount of annotated data. In this work, we improve the efficiency of the data annotation process for 3D pose estimation problems with Active Learning (AL) in a multi-view setting. AL selects examples with the highest value to annotate under limited annotation budgets (time and cost), but choosing the selection strategy is often nontrivial. We present a framework to efficiently extend existing single-view AL strategies. We then propose two novel AL strategies that make full use of multi-view geometry. Moreover, we demonstrate additional performance gains by incorporating pseudo-labels computed during the AL process, which is a form of self-training. Our system significantly outperforms simulated annotation baselines in 3D body and hand pose estimation on two large-scale benchmarks: CMU Panoptic Studio and InterHand2.6M. Notably, on CMU Panoptic Studio, we are able to reduce the turn-around time by 60% and annotation cost by 80% when compared to the conventional annotation process.

Rotation is frequently listed as a candidate for data augmentation in contrastive learning but seldom provides satisfactory improvements. We argue that this is because the rotated image is always treated as either positive or negative. The semantics of an image can be rotation-invariant or rotation-variant, so whether the rotated image is treated as positive or negative should be determined based on the content of the image. Therefore, we propose a novel augmentation strategy, adaptive Positive or Negative Data Augmentation (PNDA), in which an original and its rotated image are a positive pair if they are semantically close and a negative pair if they are semantically different. To achieve PNDA, we first determine whether rotation is positive or negative on an image-by-image basis in an unsupervised way. Then, we apply PNDA to contrastive learning frameworks. Our experiments showed that PNDA improves the performance of contrastive learning.

An autonomous driving system requires a 3D object detector, which must perceive all present road agents reliably to navigate an environment safely. However, real-world driving datasets often suffer from the problem of data imbalance, which causes difficulties in training a model that works well across all classes, resulting in an undesired imbalanced sub-optimal performance. In this work, we propose a method to address this data imbalance problem. Our method consists of two main components: (i) a LiDAR-based 3D object detector with per-class multiple detection heads where losses from each head are modified by dynamic weight average to be balanced. (ii) Contextual ground truth (GT) sampling, where we improve conventional GT sampling techniques by leveraging semantic information to augment point cloud with sampled ground truth GT objects. Our experiment with KITTI and nuScenes datasets confirms our proposed method's effectiveness in dealing with the data imbalance problem, producing better detection accuracy compared to existing approaches. Our implementation will be publicly available upon publication.

Advances in self-supervised learning, especially in contrastive learning, have drawn attention to investigating these techniques in providing effective visual representations from unlabeled images. It enables the models' ability of extracting highly consistent features by generating different views. Due to the recent success of Masked Autoencoders (MAE), an emerging trend of exploring generative modeling in self-supervised learning has come back into sight of the community. The generative approaches encode the input into a compact embedding and empower the models' ability of recovering the original input. However, in our experiments, we found vanilla MAE mainly recovers course high level semantic information and barely recovers detailed low level information. We show that in dense downstream prediction tasks like multi-organ segmentation, directly applying MAE is not ideal. In this paper, we propose RepRec, a hybrid visual representation learning framework for self-supervised pre-training on large-scale unlabelled medical datasets, which takes advantage of both contrastive and generative modeling. In our method, to solve the aforementioned dilemma that MAE encounters, a convolutional encoder is pre-trained to provide low-level feature information, in a contrastive way; and a transformer encoder is pre-trained to produce high level semantic dependency, in a generative way -- by recovering masked representations from the convolutional encoder. Extensive experiments on three multi-organ segmentation datasets demonstrate that our method outperforms current state-of-the-art methods.

Contrastive learning has shown its effectiveness in image classification and generation. Recent works apply the contrastive learning on the discriminator of the Generative Adversarial Networks, and there exists little work on exploring if contrastive learning can be applied on encoders to learn disentangled representations. In this work, we propose a simple yet effective method via incorporating contrastive learning into latent optimization, where we name it. Specifically, we first use a generator to learn discriminative and disentangled embeddings via latent optimization. Then an encoder and two momentum encoders are applied to dynamically learn disentangled information across large amount of samples with content-level and residual-level contrastive loss. In the meanwhile, we tune the encoder with the learned embeddings in an amortized manner. We evaluate our approach on ten benchmarks in terms of representation disentanglement and linear classification. Extensive experiments demonstrate the effectiveness of our ContraLORD on learning both discriminative and generative representations.

Multimodal representation learning for images with paired raw texts can improve the usability and generality of the learned semantic concepts while significantly reducing annotation costs. In this paper, we explore the design space of loss functions in visual-linguistic pretraining frameworks and propose a novel Relaxed Contrastive (ReCo) objective, which acts as a drop-in replacement of the widely used InfoNCE loss. The key insight of ReCo is to allow a relaxed negative space by not penalizing unpaired multimodal samples (ie, negative pairs) that are already orthogonal or negatively correlated. Unlike the widely-used InfoNCE, which keeps repelling negative pairs as long as they are not anti-correlated, ReCo by design embraces more diversity and flexibility of the learned embeddings. We conduct extensive experiments using ReCo with state-of-the-art models by pretraining on the MIMIC-CXR dataset that consists of chest radiographs and free-text radiology reports, and evaluating on the CheXpert dataset for multimodal retrieval and disease classification. Our ReCo achieves an absolute improvement of 2.9% over the InfoNCE baseline on the CheXpert Retrieval dataset in average retrieval precision and reports better or comparable performance in the linear evaluation and finetuning for classification. We further show that ReCo outperforms InfoNCE on the Flickr30K dataset by 1.7% in retrieval Recall@1, demonstrating the generalizability of our approach to natural images.

Object appearances change dramatically with pose variations. This creates a challenge for embedding schemes that seek to map instances with the same object ID to locations that are as close as possible. This issue becomes significantly heightened in complex computer vision tasks such as re-identification(reID). In this paper, we suggest that these dramatic appearance changes are indications that an object ID is composed of multiple natural groups, and it is counterproductive to forcefully map instances from different groups to a common location. This leads us to introduce Relation Preserving Triplet Mining (RPTM), a feature matching guided triplet mining scheme, that ensures that triplets will respect the natural subgroupings within an object ID. We use this triplet mining mechanism to establish a pose-aware, well-conditioned triplet loss by implicitly enforcing view consistency. This allows a single network to be trained with fixed parameters across datasets while providing state-of-the-art results. Code is available at https: //github.com/adhirajghosh/RPTM_reid.

Due to the scarcity of dense pixel-level semantic annotations for images recorded in adverse visual conditions, there has been a keen interest in unsupervised domain adaptation (UDA) for the semantic segmentation of such images. UDA adapts models trained on normal conditions to the target adverse-condition domains. Meanwhile, multiple datasets with driving scenes provide corresponding images of the same scenes across multiple conditions, which can serve as a form of weak supervision for domain adaptation. We propose Refign, a generic extension to self-training-based UDA methods which leverages these cross-domain correspondences. Refign consists of two steps: (1) aligning the normal-condition image to the corresponding adverse-condition image using an uncertainty-aware dense matching network, and (2) refining the adverse prediction with the normal prediction using an adaptive label correction mechanism. We design custom modules to streamline both steps and set the new state of the art for domain-adaptive semantic segmentation on several adverse-condition benchmarks, including ACDC and Dark Zurich. The approach introduces no extra training parameters, minimal computational overhead--during training only--and can be used as a drop-in extension to improve any given self-training-based UDA method. Code is available at https://github.com/brdav/refign.

In Active Domain Adaptation (ADA), one uses Active Learning (AL) to select target domain frames to annotate for Domain Adaptation (DA). Thus, ADA creates a continuum of cost-performance trade-off models, with unsupervised, and fully supervised DA techniques at the two ends. We observe that in ADA not all regions of a selected frame contribute equally to a model's performance, and there is a strong correlation between annotating certain hard/unique/novel object/stuff instances, and a model's performance. E.g., road regions in a target dataset may look mostly similar to source domain except for certain curved instances, where annotation may be more useful. Based on the observation, we propose Anchor-based and Augmentation-based ADA techniques, which, given a selected frame, determine certain 'hard' semantic regions to be annotated in that frame, such that the selected regions are complementary and diverse in the context of the current labeled set. The proposed techniques carefully avoid the pitfall of region based AL techniques which try to choose most uncertain regions in a frame, but ends up selecting all edge pixels, and similar annotation cost as the whole frame. We show that our approach achieves 66.6 \miou on \gta->\cityscapes dataset with a budget of 4.7% in comparison to 64.9 \miou by MADA [??]. Our technique can also be used as a decorator for any existing frame-based AL technique. E.g., we report 1.5% performance improvement for CDAL [??] on \cityscapes using our approach.

Most action recognition models today are highly parameterized, and evaluated on datasets with appearance-wise distinct classes. It has also been shown that 2D Convolutional Neural Networks (CNNs) tend to be biased toward texture rather than shape in still image recognition tasks, in contrast to humans. Taken together, this raises suspicion that large video models partly learn spurious spatial texture correlations rather than to track relevant shapes over time to infer generalizable semantics from their movement. A natural way to avoid parameter explosion when learning visual patterns over time is to make use of recurrence. Biological vision consists of abundant recurrent circuitry, and is superior to computer vision in terms of domain shift generalization. In this article, we empirically study whether the choice of low-level temporal modeling has consequences for texture bias and cross-domain robustness. In order to enable a light-weight and systematic assessment of the ability to capture temporal structure, not revealed from single frames, we provide the Temporal Shape (TS) dataset, as well as modified domains of Diving48 allowing for the investigation of spatial texture bias in video models. The combined results of our experiments indicate that sound physical inductive bias such as recurrence in temporal modeling may be advantageous when robustness to domain shift is important for the task.

Recent works on implicit neural representations have made significant strides. Learning implicit neural surfaces using volume rendering has gained popularity in multi-view reconstruction without 3D supervision. However, accurately recovering fine details is still challenging, due to the underlying ambiguity of geometry and appearance representation. In this paper, we present D-NeuS, a volume rendering-base neural implicit surface reconstruction method capable to recover fine geometry details, which extends NeuS by two additional loss functions targeting enhanced reconstruction quality. First, we encourage the rendered surface points from alpha compositing to have zero signed distance values, alleviating the geometry bias arising from transforming SDF to density for volume rendering. Second, we impose multi-view feature consistency on the surface points, derived by interpolating SDF zero-crossings from sampled points along rays. Extensive quantitative and qualitative results demonstrate that our method reconstructs high-accuracy surfaces with details, and outperforms the state of the art.

Most action recognition datasets and algorithms assume a closed world, where all test samples are instances of the known classes. In open set problems, test samples may be drawn from either known or unknown classes. Existing open set action recognition methods are typically based on extending closed set methods by adding post hoc analysis of classification scores or feature distances and do not capture the relations among all the video clip elements. Our approach uses the reconstruction error to determine the novelty of the video since unknown classes are harder to put back together and thus have a higher reconstruction error than videos from known classes. We refer to our solution to the open set action recognition problem as "Humpty Dumpty", due to its reconstruction abilities. Humpty Dumpty is a novel graph-based autoencoder that accounts for contextual and semantic relations among the clip pieces for improved reconstruction. A larger reconstruction error leads to an increased likelihood that the action can not be reconstructed, i.e., can not put Humpty Dumpty back together again, indicating that the action has never been seen before and is novel/unknown. Extensive experiments are performed on two publicly available action recognition datasets including HMDB-51 and UCF-101, showing the state-of-the-art performance for open set action recognition.

Procedural texts are a special type of documents that contain complex textual descriptions for carrying out a sequence of instructions. Due to the lack of visual cues, it often becomes difficult to consume the textual information effectively. In this paper, we focus on recipes - a particular type of procedural document and introduce a novel deep-learning driven system - Recipe2Video that automatically converts a recipe document into a multimodal illustrative video. Our method employs novel retrieval and re-ranking methods to select the best set of images and videos that can provide the desired illustration. We formulate a Viterbi-based optimization algorithm to stitch together a coherent video that combines the visual cues, text and voice-over to present an enhanced mode of consumption. We design automated metrics and compare performance across several baselines on two recipe datasets (RecipeQA, Tasty Videos). Our results on downstream tasks and human studies indicate that Recipe2Video captures the semantic and sequential information of the input in the generated video.

We present CoopNet, an approach that improves the cooperation of co-trained networks by dynamically adapting the apportionment of gradient, to ensure equitable learning progress. It is applied to motion-aware self-supervised prediction of depth maps, by introducing a new hybrid loss, based on a distribution model of photo-metric reconstruction errors made by, on the one hand the depth + odometry paired networks, and on the other hand the optical flow network. This model essentially assumes that the pixels from moving objects (that must be discarded for training depth and odometry), correspond to those where the two reconstructions strongly disagree. We justify this model by theoretical considerations and experimental evidences, and show that its implementation improves or is comparable to the state of the art in depth, odometry and optical flow predictions. Our code is available here: https://github.com/mhariat/CoopNet.

Recent work on image anonymization has shown that generative adversarial networks (GANs) can generate near-photorealistic faces to anonymize individuals. However, scaling up these networks to the entire human body has remained a challenging and yet unsolved task. We propose a new anonymization method that generates realistic humans for in-the-wild images. A key part of our design is to guide adversarial nets by dense pixel-to-surface correspondences between an image and a canonical 3D surface. We introduce Variational Surface-Adaptive Modulation (V-SAM) that embeds surface information throughout the generator. Combining this with our novel discriminator surface supervision loss, the generator can synthesize high quality humans with diverse appearances in complex and varying scenes. We demonstrate that surface guidance significantly improves image quality and diversity of samples, yielding a highly practical generator. Finally, we show that our method preserves data usability without infringing privacy when collecting image datasets for training computer vision models.

Low-light image enhancement has been an actively researched area for decades and has produced excellent night-time single-image, video, and Light Field restoration methods. Despite these comprehensive studies, the problem of extreme low-light stereo image enhancement has been mostly ignored. Addressing this problem can enable night-time capabilities to several applications such as smartphones and self-driving cars. We propose a light-weight and fast hybrid U-net architecture for low-light stereo image enhancement. In the initial few scale spaces, we process the left and right features individually, because the two features do not align well due to large disparity. At coarser scale-spaces, the disparity between left and right features decreases and the network's receptive field increases. We use this fact to reduce computations by simultaneously processing the left and right features, which also benefits epipole preservation. As our architecture does not use any 3D convolution for fast inference, we use an Epipole-Aware loss module to train our network. This module computes quick and coarse depth estimates to better enforce the epipolar constraints. Extensive benchmarking in terms of visual enhancement and downstream depth estimation shows that our architecture not only performs significantly better but also offers 4-60 xspeed-up with 15-100 xlower floating point operations, suitable for real-world deployment.

This paper presents a framework for real-time concealed weapon detection (CWD) on 3D radar images for walk-through screening systems. The walk-through screening system aims to ensure security in crowded areas by performing CWD on walking persons, hence it requires an accurate and real-time detection approach. To ensure accuracy, a weapon needs to be detected irrespective of its 3D orientation, thus we use the 3D radar images as detection input. For achieving real-time, we reformulate classic U-Net based segmentation networks to perform 3D detection tasks. Our 3D segmentation network predicts peak-shaped probability map, instead of voxel-wise masks, to enable position inference by elementary peak detection operation on the predicted map. In the peak-shaped probability map, the peak marks the weapon's position. So, weapon detection task translates to peak detection on the probability map. A Gaussian function is used to model weapons in the probability map. We experimentally validate our approach on realistic 3D radar images obtained from a walk-through weapon screening system prototype. Extensive ablation studies verify the effectiveness of our proposed approach over existing conventional approaches. The experimental results demonstrate that our proposed approach can perform accurate and real-time CWD, thus making it suitable for practical applications of walk-through screening.

We address the problem of face modeling, which is still challenging in achieving high-quality reconstruction results efficiently. Neither previous regression-based nor optimization-based frameworks could well balance between the facial reconstruction fidelity and efficiency. We notice that the large amount of in-the-wild facial images contain diverse appearance information, however, their underlying knowledge is not fully exploited for face modeling. To this end, we propose our Reconstruction by Encoding Facial Priors (ReEnFP) pipeline to exploit the potential of unconstrained facial images for further improvement. Our key is to encode generative priors learned by a style-based texture generator on unconstrained data for fast and detail-preserving face reconstruction. With our texture generator pre-trained using a differentiable renderer, faces could be encoded to its latent space as opposed to the time-consuming optimization-based inversion. Our generative prior encoding is further enhanced with a pyramid fusion block for adaptive integration of input spatial information. Extensive experiments show that our method reconstructs photo-realistic facial textures and geometric details with precise identity recovery.