WACV 2023 Papers

Deep Learning-based 2D/3D registration enables fast, robust, and accurate X-ray to CT image fusion when large annotated paired datasets are available for training. However, the need for paired CT volume and X-ray images with ground truth registration limits the applicability in interventional scenarios. An alternative is to use simulated X-ray projections from CT volumes, thus removing the need for paired annotated datasets. Deep Neural Networks trained exclusively on simulated X-ray projections can perform significantly worse on real X-ray images due to the domain gap. We propose a self-supervised 2D/3D registration framework combining simulated training with unsupervised feature and pixel space domain adaptation to overcome the domain gap and eliminate the need for paired annotated datasets. Our framework achieves a registration accuracy of 1.83 +-1.16 mm with a high success ratio of 90.1% on real X-ray images showing a 23.9% increase in success ratio compared to reference annotation-free algorithms.

For change detection in remote sensing, constructing a training dataset for deep learning models is quite difficult due to the requirements of bi-temporal supervision. To overcome this issue, single-temporal supervision which treats change labels as the difference of two semantic masks has been proposed. This novel method trains a change detector using two spatially unrelated images with corresponding semantic labels. However, training with unpaired dataset shows not enough performance compared with other methods based on bi-temporal supervision. We suspect this phenomenon caused by ignorance of meaningful information in the actual bi-temporal pairs.In this paper, we emphasize that the change originates from the source image and show that manipulating the source image as an after-image is crucial to the performance of change detection. Our method achieves state-of-the-art performance in a large gap than existing methods.

We present a new method for lightweight novel-view synthesis that generalizes to an arbitrary forward-facing scene. Recent approaches are computationally expensive, require per-scene optimization, or produce a memory-expensive representation. We start by representing the scene with a set of fronto-parallel semitransparent planes and afterwards convert them to deformable layers in an end-to-end manner. Additionally, we employ a feed-forward refinement procedure that corrects the estimated representation by aggregating information from input views. Our method does not require any fine-tuning when a new scene is processed and can handle an arbitrary number of views without any restrictions. Experimental results show that our approach surpasses recent models in terms of both common metrics and human evaluation, with the noticeable advantage in inference speed and compactness of the inferred layered geometry.

State-of-the-art frameworks in self-supervised learning have recently shown that fully utilizing transformer-based models can lead to performance boost compared to conventional CNN models. Striving to maximize the mutual information of two views of an image, existing works apply a contrastive loss to the final representations. Motivated by self-distillation in the supervised regime, we further exploit this by allowing the intermediate representations to learn from the final layer via the contrastive loss. Through self-distillation, the intermediate layers are better suited for instance discrimination, making the performance of an early-exited sub-network not much degraded from that of the full network. This renders the pretext task easier also for the final layer, lead to better representations. Our method, Self-Distilled Self-Supervised Learning (SDSSL), outperforms competitive baselines (SimCLR, BYOL and MoCo v3) using ViT on various tasks and datasets. In the linear evaluation and k-NN protocol, SDSSL not only leads to superior performance in the final layers, but also in most of the lower layers. Furthermore, qualitative and quantative analyses show how representations are formed more effectively along the transformer layers. Code will be available.

Point cloud classification is a popular task in 3D vision. However, previous works, usually assume that point clouds at test time are obtained with the same procedure or sensor as those at training time. Unsupervised Domain Adaptation (UDA) instead, breaks this assumption and tries to solve the task on an unlabeled target domain, leveraging only on a supervised source domain. For point cloud classification, recent UDA methods try to align features across domains via auxiliary tasks such as point cloud reconstruction, which however do not optimize the discriminative power in the target domain in feature space. In contrast, in this work, we focus on obtaining a discriminative feature space for the target domain enforcing consistency between a point cloud and its augmented version. We then propose a novel iterative self-training methodology that exploits Graph Neural Networks in the UDA context to refine pseudo-labels. We perform extensive experiments and set the new state-of-the art in standard UDA benchmarks for point cloud classification. Finally, we show how our approach can be extended to more complex tasks such as part segmentation.

Efficient custom pooling techniques that can aggressively trim the dimensions of a feature map for resource-constrained computer vision applications have recently gained significant traction. However, prior pooling works extract only the local context of the activation maps, limiting their effectiveness. In contrast, we propose a novel non-local self-attentive pooling method that can be used as a drop-in replacement to the standard pooling layers, such as max/average pooling or strided convolution. The proposed self-attention module uses patch embedding, multi-head self-attention, and spatial-channel restoration, followed by sigmoid activation and exponential soft-max. This self-attention mechanism efficiently aggregates dependencies between non-local activation patches during down-sampling. Extensive experiments on standard object classification and detection tasks with various convolutional neural network (CNN) architectures demonstrate the superiority of our proposed mechanism over the state-of-the-art (SOTA) pooling techniques. In particular, we surpass the test accuracy of existing pooling techniques on different variants of MobileNet-V2 on ImageNet by an average of 1.2%. With the aggressive down-sampling of the activation maps in the initial layers (providing up to 22x reduction in memory consumption), our approach achieves 1.43% higher test accuracy compared to SOTA techniques with iso-memory footprints. This enables the deployment of our models in memory-constrained devices, such as micro-controllers without losing significant accuracy, because the initial activation maps consume a significant amount of on-chip memory for high-resolution images required for complex vision tasks. Our pooling method also leverages channel pruning to further reduce memory footprints. Codes are available at https://github.com/C-Fun/Non-Local-Pooling.

Humans have a unique ability to learn new representations from just a handful of examples with little to no supervision. Deep learning models, however, require an abundance of data and supervision to perform at a satisfactory level. Unsupervised few-shot learning (U-FSL) is the pursuit of bridging this gap between machines and humans. Inspired by the capacity of graph neural networks (GNNs) in discovering complex inter-sample relationships, we propose a novel self-attention based message passing contrastive learning approach (coined as SAMP-CLR) for U-FSL pre-training. We also propose an optimal transport (OT) based fine-tuning strategy (we call OpT-Tune) to efficiently induce task awareness into our novel end-to-end unsupervised few-shot classification framework (SAMPTransfer). Our extensive experimental results corroborate the efficacy of SAMPTransfer in a variety of downstream few-shot classification scenarios, setting a new state-of-the-art for U-FSL on both miniImageNet and tieredImageNet benchmarks, offering up to 7%+ and 5%+ improvements, respectively. Our further investigations also confirm that SAMPTransfer remains on-par with some supervised baselines on miniImageNet and outperforms all existing U-FSL baselines in a challenging cross-domain scenario.

The resurgence of deep neural networks has created an alternative pathway for low-dose computed tomography denoising by learning a nonlinear transformation function between low-dose CT (LDCT) and normal-dose CT (NDCT) image pairs. However, those paired LDCT and NDCT images are rarely available in the clinical environment, making deep neural network deployment infeasible. This study proposes a novel method for self-supervised low-dose CT denoising to alleviate the requirement of paired LDCT and NDCT images. Specifically, we have trained an invertible neural network to minimize the pixel-based mean square distance between a noisy slice and the average of its two immediate adjacent noisy slices. We have shown the aforementioned is similar to training a neural network to minimize the distance between clean NDCT and noisy LDCT image pairs. Again, during the reverse mapping of the invertible network, the output image is mapped to the original input image, similar to cycle consistency loss. Finally, the trained invertible network's forward mapping is used for denoising LDCT images. Extensive experiments on two publicly available datasets showed that our method performs favourably against other existing unsupervised methods.

Partial domain adaptation which assumes that the unknown target label space is a subset of the source label space has attracted much attention in computer vision. Despite recent progress, existing methods often suffer from three key problems: negative transfer, lack of discriminability, and domain invariance in the latent space. To alleviate the above issues, we develop a novel 'Select, Label, and Mix' (SLM) framework that aims to learn discriminative invariant feature representations for partial domain adaptation. First, we present an efficient "select" module that automatically filters out the outlier source samples to avoid negative transfer while aligning distributions across both domains. Second, the "label" module iteratively trains the classifier using both the labeled source domain data and the generated pseudo-labels for the target domain to enhance the discriminability of the latent space. Finally, the "mix" module utilizes domain mixup regularization jointly with the other two modules to explore more intrinsic structures across domains leading to a domain-invariant latent space for partial domain adaptation. Extensive experiments on several benchmark datasets demonstrate the superiority of our proposed framework over state-of-the-art methods. Project page: https://cvir.github.io/projects/slm.

This paper proposes an efficient iris localization method without using iris segmentation and circle fitting. Conventional iris localization methods first extract iris regions by using semantic segmentation methods such as U-Net. Afterward, the inner and outer iris circles are localized using the traditional circle fitting algorithm. However, this approach requires high-resolution encoder-decoder networks for iris segmentation, so it causes computational costs to be high. In addition, traditional circle fitting tends to be sensitive to noise in input images and fitting parameters, causing the iris recognition performance to be poor. To solve these problems, we propose an iris localization network (ILN), that can directly localize pupil and iris circles with eyelid points from a low-resolution iris image. We also introduce a pupil refinement network (PRN) to improve the accuracy of pupil localization. Experimental results show that the combination of ILN and PRN works in 34.5 ms for one iris image on a CPU, and its localization performance outperforms conventional iris segmentation methods. In addition, generalized evaluation results show that the proposed method has higher robustness for datasets in different domain than other segmentation methods. Furthermore, we also confirm that the proposed ILN and PRN improve the iris recognition accuracy.

We propose Seg&Struct, a supervised learning framework leveraging the interplay between part segmentation and structure inference and demonstrating their synergy in an integrated framework. Both part segmentation and structure inference have been extensively studied in the recent deep learning literature, while the supervisions used for each task have not been fully exploited to assist the other task. Namely, structure inference has been typically conducted with an autoencoder that does not leverage the point-to-part associations. Also, segmentation has been mostly performed without structural priors that tell the plausibility of the output segments. We present how these two tasks can be best combined while fully utilizing supervision to improve performance. Our framework first decomposes a raw input shape into part segments using an off-the-shelf algorithm, whose outputs are then mapped to nodes in a part hierarchy, establishing point-to-part associations. Following this, ours predicts the structural information, e.g., part bounding boxes and part relationships. Lastly, the segmentation is rectified by examining the confusion of part boundaries using the structure-based part features. Our experimental results based on the StructureNet and PartNet demonstrate that the interplay between the two tasks results in remarkable improvements in both tasks: 27.91% in structure inference and 0.5% in segmentation.

Binary neural networks (BNNs) are advantageous in performance and memory footprint but suffer from low accuracy due to their limited expression capability. Recent works have tried to enhance the accuracy of BNNs via a gradient-based search algorithm and showed promising results. However, the mixture of architecture search and binarization induce the instability of the search process, resulting in convergence to the suboptimal point. To address this issue, we propose a BNN architecture search framework with bimodal parameter perturbation. The bimodal parameter perturbation can improve the stability of gradient-based architecture search by reducing the sharpness of the loss surface along both weight and architecture parameter axes. In addition, we refine the inverted bottleneck convolution block for having robustness with BNNs. The synergy of the refined space and the stabilized search process allows us to find out the accurate BNNs with high computation efficiency. Experimental results show that our framework finds the best architecture on CIFAR-100 and ImageNet datasets in the existing search space for BNNs. We also tested our framework on another search space based on the inverted bottleneck convolution block, and the selected BNN models using our approach achieved the highest accuracy on both datasets with a much smaller number of equivalent operations than previous works.

For aiming at a more accurate appearance-based gaze estimation, a series of recent works propose to use transformers or high-resolution networks in several ways which achieve state-of-the-art, but such works lack efficiency for real-time applications on edge computing devices. In this paper, we propose a compact model to precisely and efficiently solve gaze estimation. The proposed model includes 1) a Neural Architecture Search(NAS)-based multi-resolution feature extractor for extracting feature maps with global and local information which are essential for this task and 2) a novel multi-resolution fusion transformer as the gaze estimation head for efficiently estimating gaze values by fusing the extracted feature maps. We search our proposed model, called GazeNAS-ETH, on the ETH-XGaze dataset. We confirmed through experiments that GazeNAS-ETH achieved state-of-the-art on Gaze360, MPIIFaceGaze, RTGENE, and EYEDIAP datasets, while having only about 1M parameters and using only 0.28 GFLOPs, which is significantly less compared to previous state-of-the-art models, making it easier to deploy for real-time applications.

Recent years have witnessed the success of deep learning on the visual sound separation task. However, existing works follow similar settings where the training and testing datasets share the same musical instrument categories, which to some extent limits the versatility of this task. In this work, we focus on a more general and challenging scenario, namely the separation of unknown musical instruments, where the categories in training and testing phases have no overlap with each other. To tackle this new setting, we propose the "Separation-with-Consistency" (SeCo) framework, which can accomplish the separation on unknown categories by exploiting the consistency constraints. Furthermore, to capture richer characteristics of the novel melodies, we devise an online matching strategy, which can bring stable enhancements with no cost of extra parameters. Experiments demonstrate that our SeCo framework exhibits strong adaptation ability on the novel musical categories and outperforms the baseline methods by a notable margin.

Progress in digital pathology is hindered by high-resolution images and the prohibitive cost of exhaustive localized annotations. The commonly used paradigm to categorize pathology images is patch-based processing, which often incorporates multiple instance learning MIL to aggregate local patch-level representations yielding image-level prediction. Nonetheless, diagnostically relevant regions may only take a small fraction of the whole tissue, and current MIL-based approaches often process images uniformly, discarding the inter-patches interactions. To alleviate these issues, we propose ScoreNet, a new efficient transformer that exploits a differentiable recommendation stage to extract discriminative image regions and dedicate computational resources accordingly. The proposed transformer leverages the local and global attention of a few dynamically recommended high-resolution regions at an efficient computational cost. We further introduce a novel mixing data-augmentation, namely ScoreMix, by leveraging the image's semantic distribution to guide the data mixing and produce coherent sample-label pairs. ScoreMix is embarrassingly simple and mitigates the pitfalls of previous augmentations, which assume a uniform semantic distribution and risk mislabeling the samples. Thorough experiments and ablation studies on three breast cancer histology datasets of Haematoxylin & Eosin (H&E) have validated the superiority of our approach over prior arts, including transformer-based models on tumour regions-of-interest TRoIs classification. ScoreNet equipped with proposed ScoreMix augmentation demonstrates better generalization capabilities and achieves new state-of-the-art (SOTA) results with only 50% of the data compared to other mixing augmentation variants. Finally, ScoreNet yields high efficacy and outperforms SOTA efficient transformers, namely TransPath and SwinTransformer, with throughput around 3x and 4x higher than the aforementioned architectures, respectively.

In this paper, we propose the first-ever real benchmark thought for evaluating Neural Radiance Fields (NeRFs) and, in general, Neural Rendering (NR) frameworks. We design and implement an effective pipeline for scanning real objects in quantity and effortlessly. Our scan station is built with less than 500 hardware budget and can collect roughly 4000 images of a scanned object in just 5 minutes. Such a platform is used to build ScanNeRF, a dataset characterized by several train/val/test splits aimed at benchmarking the performance of modern NeRF methods under different conditions. Accordingly, we evaluate three cutting-edge NeRF variants on it to highlight their strengths and weaknesses. The dataset is available on our project page, together with an online benchmark to foster the development of better and better NeRFs.

A critical object detection task is finetuning an existing model to detect novel objects, but the standard workflow requires bounding box annotations which are time-consuming and expensive to collect. Weakly supervised object detection (WSOD) offers an appealing alternative, where object detectors can be trained using image-level labels. However, the practical application of current WSOD models is limited, as they only operate at small data scales and require multiple rounds of training and refinement. To address this, we propose the Weakly Supervised Detection Transformer, which enables efficient knowledge transfer from a large-scale pretraining dataset to WSOD finetuning on hundreds of novel objects. Additionally, we leverage pretrained knowledge to improve the multiple instance learning (MIL) framework often used in WSOD methods. Our experiments show that our approach outperforms previous state-of-the-art models on large-scale novel object detection datasets, and our scaling study reveals that class quantity is more important than image quantity for WSOD pretraining.

Recent neural rendering approaches greatly improve image quality, reaching near photorealism. However, the underlying neural networks have high runtime, precluding telepresence and virtual reality applications that require high resolution at low latency. The sequential dependency of layers in deep networks makes their optimization difficult. We break this dependency by caching information from the previous frame to speed up the processing of the current one with an implicit warp. The warping with a shallow network reduces latency and the caching operations can further be parallelized to improve the frame rate. In contrast to existing temporal neural networks, ours is tailored for the task of rendering novel views of faces by conditioning on the change of the underlying surface mesh. We test the approach on view-dependent rendering of 3D portrait avatars, as needed for telepresence, on established benchmark sequences. Warping reduces latency by 70% (from 49.4ms to 14.9ms on commodity GPUs) and scales frame rates accordingly over multiple GPUs while reducing image quality by only 1%, making it suitable as part of end-to-end view-dependent 3D teleconferencing applications.

Artificial learning systems aspire to mimic human intelligence by continually learning from a stream of tasks without forgetting past knowledge. One way to enable such learning is to store past experiences in the form of input examples in episodic memory and replay them when learning new tasks. However, performance of such method suffers as the size of the memory becomes smaller. In this paper, we propose a new approach for experience replay, where we select the past experiences by looking at the saliency maps which provide visual explanations for the model's decision. Guided by these saliency maps, we pack the memory with only the parts or patches of the input images important for the model's prediction. While learning a new task, we replay these memory patches with appropriate zero-padding to remind the model about its past decisions. We evaluate our algorithm on CIFAR-100, miniImageNet and CUB datasets and report better performance than the state-of-the-art approaches. With qualitative and quantitative analyses we show that our method captures richer summaries of past experiences without any memory increase, and hence performs well with small episodic memory.