WACV 2024 论文列表

Single-photon avalanche diodes (SPADs) are detectors capable of capturing single photons and of performing photon counting. SPADs have an exceptional temporal resolution and are thus highly suitable for time-resolved imaging applications. Applications span from biomedical research to consumers with SPADs integrated in smartphones and mixed-reality headsets. While conventional SPAD imaging systems typically employ photon time-tagging and histogram-building in the workflow, the pulse signal output of a SPAD naturally lends itself as input to spiking neural networks (SNNs). Leveraging this potential, SNNs offer real-time, energy-efficient, and intelligent processing with high throughput. In this paper, we propose two SNN frameworks, namely the Transporter SNN and the Reversed Start-stop SNN, along with corresponding hardware schemes for active time-resolved SPAD imaging. These frameworks convert phase-coded spike trains into density- and interspike-interval-coded ones, enabling training with rate-based warm-up and Surrogate Gradient. The SNNs are evaluated on fluorescence lifetime imaging. The results demonstrate that the accuracy of shallow SNNs is on par with established benchmarks. Our vision is to integrate SNNs in SPAD sensors and to explore advanced SNNs within the proposed schemes for high-level applications.

Spiking neural networks (SNNs) have ultra-low energy consumption and high biological plausibility due to their binary and bio-driven nature compared with artificial neural networks (ANNs). While previous research has primarily focused on enhancing the performance of SNNs in classification tasks, the generative potential of SNNs remains relatively unexplored. In our paper, we put forward Spiking Denoising Diffusion Probabilistic Models (SDDPM), a new class of SNN-based generative models that achieve high sample quality. To fully exploit the energy efficiency of SNNs, we propose a purely Spiking U-Net architecture, which achieves comparable performance to its ANN counterpart using only 4 time steps, resulting in significantly reduced energy consumption. Extensive experimental results reveal that our approach achieves state-of-the-art on the generative tasks and substantially outperforms other SNN-based generative models, achieving up to 12x and 6x improvement on the CIFAR-10 and the CelebA datasets, respectively. Moreover, we propose a threshold-guided strategy that can further improve the performances by 2.69% in a training-free manner. The SDDPM symbolizes a significant advancement in the field of SNN generation, injecting new perspectives and potential avenues of exploration. Our code is available at https://github.com/AndyCao1125/SDDPM.

This paper introduces SphereCraft, a dataset specifically designed for spherical keypoint detection, matching, and camera pose estimation. The dataset addresses the limitations of existing datasets by providing extracted keypoints from various detectors, along with their ground truth correspondences. Synthetic scenes with photo-realistic rendering and accurate 3D meshes are included, as well as real-world scenes acquired from different spherical cameras. SphereCraft enables the development and evaluation of algorithms targeting multiple camera viewpoints, advancing the state-of-the-art in computer vision tasks involving spherical images. Our dataset is available at https://dfki.github.io/spherecraftweb/.

This paper addresses the problem of reconstructing scenes behind optical diffusers, which is common in applications such as imaging through frosted glass. We propose a new approach that exploits specular reflection to capture sharp light distributions with a point light source, which can be used to detect reflections in low signal-to-noise scenarios. In this paper, we propose a rasterizer-based differentiable renderer to solve this problem by minimizing the difference between the captured and rendered images. Because our method can simultaneously optimize multiple observations for different light source positions, it is confirmed that ambiguities of the scene are efficiently eliminated by increasing the number of observations. Experiments show that the proposed method can reconstruct a scene with several mirror-like objects behind the diffuser in both simulated and real environments.

Underwater images often suffer from color distortion, haze, and limited visibility due to light refraction and absorption in water. These challenges significantly impact autonomous underwater vehicle applications, necessitating efficient image enhancement techniques. To address these challenges, we propose a Multi-Domain Query Cascaded Transformer Network for underwater image enhancement. Our approach includes a novel Multi-Domain Query Cascaded Attention mechanism that integrates localized transmission features and global illumination features. To improve feature propagation from the encoder to the decoder, we propose a Spatio-Spectro Fusion-Based Attention Block. Additionally, we introduce a Hybrid Fourier-Spatial Upsampling Block, which uniquely combines Fourier and spatial upsampling techniques to enhance feature resolution effectively. We evaluate our method on benchmark synthetic and real-world underwater image datasets, demonstrating its superiority through extensive ablation studies and comparative analysis. The testing code is available at: https: //github.com/Mdraqibkhan/Spectroformer.

Owing to the power of vision-language foundation models, e.g., CLIP, the area of image synthesis has seen recent important advances. Particularly, for style transfer, CLIP enables transferring more general and abstract styles without collecting the style images in advance, as the style can be efficiently described with natural language, and the result is optimized by minimizing the CLIP similarity between the text description and the stylized image. However, directly using CLIP to guide style transfer leads to undesirable artifacts (mainly written words and unrelated visual entities) spread over the image. In this paper, we propose SpectralCLIP, which is based on a spectral representation of the CLIP embedding sequence, where most of the common artifacts occupy specific frequencies. By masking the band including these frequencies, we can condition the generation process to adhere to the target style properties (e.g., color, texture, paint stroke, etc.) while excluding the generation of larger-scale structures corresponding to the artifacts. Experimental results show that SpectralCLIP prevents the generation of artifacts effectively in quantitative and qualitative terms, without impairing the stylisation quality. We also apply SpectralCLIP to text-conditioned image generation and show that it prevents written words in the generated images. Our code is available at https://github.com/zipengxuc/SpectralCLIP.

As 2D-CNNs are growing in image recognition literature, 3D-CNNs are enthusiastically applied to video action recognition. While spatio-temporal (3D) convolution successfully stems from spatial (2D) convolution, it is still unclear how the convolution works for encoding temporal motion patterns in 3D-CNNs. In this paper, we shed light on the mechanism of feature extraction through analyzing the spatio-temporal filters from a temporal viewpoint. The analysis not only describes characteristics of the two action datasets, Something-Something-v2 (SSv2) and Kinetics-400, but also reveals how temporal dynamics are characterized through stacked spatio-temporal convolutions. Based on the analysis, we propose methods to improve temporal feature extraction, covering temporal filter representation and temporal data augmentation. The proposed method contributes to enlarging temporal receptive field of 3D-CNN without touching its fundamental architecture, thus keeping the computation cost. In the experiments on action classification using SSv2 and Kinetics-400, it produces favorable performance improvement of 3D-CNNs.

Reconstructing accurate 3D surfaces from noisy point clouds is a fundamental problem in computer vision. Among different approaches, neural implicit methods that map 3D coordinates to occupancy values benefit from the learning capabilities of deep neural networks and the flexible topology of implicit representations, and achieve promising reconstruction results. However, existing methods utilize standard (dense) 3D convolutional neural networks for feature extraction and occupancy prediction, which significantly restricts the capability to reconstruct details. In this paper, we propose a neural implicit method based on sparse convolutions, where features and network calculations only focus on grid points close to the surface to be reconstructed. This allows us to build significantly higher resolution 3D grids and reconstruct high-fidelity details. We further build a 3D residual UNet to extract features which are robust to noise, while ensuring details are retained. A 3D position along with features extracted at the position are fed into the occupancy probability predictor network to obtain occupancy. As features at nearby grid points to the query position may not exist due to the sparse nature, we propose a normalized weight interpolation approach to obtain smooth interpolation with sparse data. Experimental results demonstrate that our method achieves promising results, both qualitatively and quantitatively, outperforming existing methods.

Online Unsupervised Domain Adaptation (OUDA) for person Re-Identification (Re-ID) is the task of continuously adapting a model trained on a well-annotated source-domain dataset to a target domain observed as a data stream. In OUDA, person Re-ID models face two main challenges: catastrophic forgetting and domain shift. In this work, we propose a new Source-guided Similarity Preservation (S2P) framework to alleviate these two problems. Our framework is based on the extraction of a support set composed of source images that maximizes the similarity with the target data. This support set is used to identify feature similarities that must be preserved during the learning process. S2P can incorporate multiple existing UDA methods to mitigate catastrophic forgetting. Our experiments show that S2P outperforms previous state-of-the-art methods on multiple real-to-real and synthetic-to-real challenging OUDA benchmarks.

Spatial localization of 3D sound sources is an important problem in many real world scenarios, especially when the sources may not have any visually distinguishable characteristics; e.g., finding a gas leak, a malfunctioning motor, etc. In this paper, we cast this task in a novel audio-visual setting, by introducing an acoustic-camera rig consisting of a centered pinhole RGB camera and an uniform circular array of four coplanar microphones. Using this setup, we propose Sound3DVDet - a 3D sound source localization Transformer model that takes as input the neural embeddings of the sound signals from the microphones and multiview images (with known poses), and learns to minimize the reprojection error between the predicted locations of the sound sources by the two modalities and the ground truth as the camera moves. When training to minimize this consistency loss, the model learns an implicit association between the audio heard at the microphones and the 3D spatial location in the RGB image, which is sufficient to localize the sources in 3D from a single RGB view. To evaluate our method, we introduce a new dataset: Sound3DVDet Dataset, consisting of nearly 6k scenes produced using the SoundSpaces simulator. We conduct extensive experiments on our dataset and shows the efficacy of our approach against closely related methods, demonstrating significant improvements in the localization accuracy.

Drone-based land surveys and tracking applications with a moving camera require three-dimensional reconstructions from videos recorded using a downward facing camera and are usually generated by Structure-from-Motion (SfM) algorithms. Unfortunately, monocular SfM pipelines can fail in the presence of lens distortion due to a critical configuration resulting in a plane-sphere ambiguity which is characterized by severe curvatures of the reconstructions and erroneous relative camera pose estimations. We propose a 4-point minimal solver for the relative pose estimation for two views sharing the same radial distortion parameters (i.e. from the same camera) with a viewing direction perpendicular to the ground plane. To extract 3D reconstructions from continuous videos, the relative pose of pairwise frames is estimated by using the solver with RANSAC and the Sampson error where globally consistent distortion parameters are determined by taking the medial of all values. Moreover, we propose an additional regularizer for the final bundle adjustment to remove any remaining curvature of the reconstruction if necessary. We tested our methods on synthetic and real-world data and our results demonstrate a significant reduction of curvature and more accurate relative pose estimations. Our algorithm can be easily integrated into existing pipelines and is therefore a practical solution to resolve the plane-sphere ambiguity in a variety of top-down SfM applications.

Label-noise or curated unlabeled data are used to compensate for the assumption of clean labeled data in training the conditional generative adversarial network; however, satisfying such an extended assumption is occasionally laborious or impractical. As a step towards generative modeling accessible to everyone, we introduce a novel conditional image generation framework that accepts noisy-labeled and uncurated unlabeled data during training: (i) closed-set and open-set label noise in labeled data and (ii) closed-set and open-set unlabeled data. To combat it, we propose soft curriculum learning, which assigns instance-wise weights for adversarial training while assigning new labels for unlabeled data and correcting wrong labels for labeled data. Unlike popular curriculum learning, which uses a threshold to pick the training samples, our soft curriculum controls the effect of each training instance by using the weights predicted by the auxiliary classifier, resulting in the preservation of useful samples while ignoring harmful ones. Our experiments show that our approach outperforms existing semi-supervised and label-noise robust methods in terms of both quantitative and qualitative performance. In particular, the proposed approach matches the performance of (semi-)supervised GANs even with less than half the labeled data.

This work introduces a multi-camera tracking dataset consisting of 234 hours of video data recorded concurrently from 234 overlapping HD cameras covering a 4.2 mile stretch of 8-10 lane interstate highway near Nashville, TN. The video is recorded during a period of high traffic density with 500+ objects typically visible within the scene and typical object longevities of 3-15 minutes. GPS trajectories from 270 vehicle passes through the scene are manually corrected in the video data to provide a set of ground-truth trajectories for recall-oriented tracking metrics, and object detections are provided for each camera in the scene (159 million total before cross-camera fusion). Initial benchmarking of tracking-by-detection algorithms is performed against the GPS trajectories, and a best HOTA of only 9.5% is obtained (best recall 75.9% at IOU 0.1, 47.9 average IDs per ground truth object), indicating the benchmarked trackers do not perform sufficiently well at the long temporal and spatial durations required for traffic scene understanding.

Weakly-supervised semantic segmentation (WSSS) performs pixel-wise classification given only image-level labels for training. Despite the difficulty of this task, the research community has achieved promising results over the last five years. Still, current WSSS literature misses the detailed sense of how well the methods perform on different sizes of objects. Thus we propose a novel evaluation metric to provide a comprehensive assessment across different object sizes and collect a size-balanced evaluation set to complement PASCAL VOC. With these two gadgets, we reveal that the existing WSSS methods struggle in capturing small objects. Furthermore, we propose a size-balanced cross-entropy loss coupled with a proper training strategy. It generally improves existing WSSS methods as validated upon ten baselines on three different datasets.

Interactive segmentation methods have been investigated to address the potential need for additional refinement in automatic segmentation via human-in-the-loop techniques. For accurate segmentation of 3D images, we propose Slice-and-Conquer, a novel planar-to-3D pipeline formulating volumetric mask construction into two stages: 1) 2D interactive segmentation and 2) guided 3D segmentation. Specifically, the first stage enables users to focus on a single 2D slice and provides the corresponding 2D prediction results as strong shape priors. Taking the planar guidance, an accurate 3D mask can be constructed with minimal interactions. To support a flexible iterative refinement, our system recommends a next slice to annotate at the end of the second stage. Since volumetric segmentation can be completed by consecutively annotating a few recommended 2D slices, our method significantly reduces the cognitive burden of exploring volumetric space for users. Through extensive experiments on various datasets of 3D biomedical images, we demonstrate the effectiveness of the proposed pipeline.

We introduce Sketch-based Video Object Localization (SVOL), a new task aimed at localizing spatio-temporal object boxes in video queried by the input sketch. We first outline the challenges in the SVOL task and build the Sketch-Video Attention Network (SVANet) with the following design principles: (i) to consider temporal information of video and bridge the domain gap between sketch and video; (ii) to accurately identify and localize multiple objects simultaneously; (iii) to handle various styles of sketches; (iv) to be classification-free. In particular, SVANet is equipped with a Cross-modal Transformer that models the interaction between learnable object tokens, query sketch, and video through attention operations, and learns upon a per-frame set matching strategy that enables frame-wise prediction while utilizing global video context. We evaluate SVANet on a newly curated SVOL dataset. By design, SVANet successfully learns the mapping between the query sketches and video objects, achieving state-of-the-art results on the SVOL benchmark. We further confirm the effectiveness of SVANet via extensive ablation studies and visualizations. Lastly, we demonstrate its transfer capability on unseen datasets and novel categories, suggesting its high scalability in real-world applications. Codes are available at https://github.com/sangminwoo/SVOL.

Blurry appearance of fast moving objects in video frames was successfully used to reconstruct the object appearance and motion in both 2D and 3D domains. The proposed method addresses the novel, severely ill-posed, task of single-image fast moving object deblurring, shape, and trajectory recovery -- previous approaches require at least three consecutive video frames. Given a single image, the method outputs the object 2D appearance and position in a series of sub-frames as if captured by a high-speed camera (i.e. temporal super-resolution). The proposed SI-DDPM-FMO method is trained end-to-end on a synthetic dataset with various moving objects, yet it generalizes well to real-world data from several publicly available datasets. SI-DDPM-FMO performs similarly to or better than recent multi-frame methods and a carefully designed baseline method.

In recent years, the research community has shown a lot of interest to panoramic images that offer a 360-degree directional perspective. Multiple data modalities can be fed, and complimentary characteristics can be utilized for more robust and rich scene interpretation based on semantic segmentation, to fully realize the potential. Existing research, however, mostly concentrated on pinhole RGB-X semantic segmentation. In this study, we propose a transformer-based cross-modal fusion architecture to bridge the gap between multi-modal fusion and omnidirectional scene perception. We employ distortion-aware modules to address extreme object deformations and panorama distortions that result from equirectangular representation. Additionally, we conduct cross-modal interactions for feature rectification and information exchange before merging the features in order to communicate long-range contexts for bi-modal and tri-modal feature streams. In thorough tests using combinations of four different modality types in three indoor panoramic-view datasets, our technique achieved state-of-the-art mIoU performance: 60.60% on Stanford2D3DS (RGB-HHA), 71.97% on Structured3D (RGB-D-N), and 35.92% on Matterport3D (RGB-D).

Deep learning techniques often perform poorly in the presence of domain shift, where the test data follows a different distribution than the training data. The most practically desirable approach to address this issue is Single Domain Generalization (S-DG), which aims to train robust models using data from a single source. Prior work on S-DG has primarily focused on using data augmentation techniques to generate diverse training data. In this paper, we explore an alternative approach by investigating the robustness of linear operators, such as convolution and dense layers commonly used in deep learning. We propose a novel operator called XCNorm that computes the normalized cross-correlation between weights and an input feature patch. This approach is invariant to both affine shifts and changes in energy within a local feature patch and eliminates the need for commonly used non-linear activation functions. We show that deep neural networks composed of this operator are robust to common semantic distribution shifts. Furthermore, our empirical results on single-domain generalization benchmarks demonstrate that our proposed technique performs comparably to the state-of-the-art methods.