ICCV 2023 Papers

Sparse instance-level supervision has recently been explored to address insufficient annotation in biomedical instance segmentation, which is easier to annotate crowded instances and better preserves instance completeness for 3D volumetric datasets compared to common semi-supervision.In this paper, we propose a sparsely supervised biomedical instance segmentation framework via cross-representation affinity consistency regularization. Specifically, we adopt two individual networks to enforce the perturbation consistency between an explicit affinity map and an implicit affinity map to capture both feature-level instance discrimination and pixel-level instance boundary structure. We then select the highly confident region of each affinity map as the pseudo label to supervise the other one for affinity consistency learning. To obtain the highly confident region, we propose a pseudo-label noise filtering scheme by integrating two entropy-based decision strategies. Extensive experiments on four biomedical datasets with sparse instance annotations show the state-of-the-art performance of our proposed framework. For the first time, we demonstrate the superiority of sparse instance-level supervision on 3D volumetric datasets, compared to common semi-supervision under the same annotation cost.

Referring video object segmentation (RVOS), as a supervised learning task, relies on sufficient annotated data for a given scene. However, in more realistic scenarios, only minimal annotations are available for a new scene, which poses significant challenges to existing RVOS methods. With this in mind, we propose a simple yet effective model with a newly designed cross-modal affinity (CMA) module based on a Transformer architecture. The CMA module builds multimodal affinity with a few samples, thus quickly learning new semantic information, and enabling the model to adapt to different scenarios. Since the proposed method targets limited samples for new scenes, we generalize the problem as - few-shot referring video object segmentation (FS-RVOS). To foster research in this direction, we build up a new FS-RVOS benchmark based on currently available datasets. The benchmark covers a wide range and includes multiple situations, which can maximally simulate real-world scenarios. Extensive experiments show that our model adapts well to different scenarios with only a few samples, reaching state-of-the-art performance on the benchmark. On Mini-Ref-YouTube-VOS, our model achieves an average performance of 53.1 J and 54.8 F, which are 10% better than the baselines. Furthermore, we show impressive results of 77.7 J and 74.8 F on Mini-Ref-SAIL-VOS, which are significantly better than the baselines. Code is publicly available at https://github.com/hengliusky/Few_shot_RVOS.

Although existing image deep learning super-resolution (SR) methods achieve promising performance on benchmark datasets, they still suffer from severe performance drops when the degradation of the low-resolution (LR) input is not covered in training. To address the problem, we propose an innovative unsupervised method of Learning Correction Filter via Degradation-Adaptive Regression for Blind Single Image Super-Resolution. Highly inspired by the generalized sampling theory, our method aims to enhance the strength of off-the-shelf SR methods trained on known degradations and adapt to unknown complex degradations to generate improved results. Specifically, we first conduct degradation estimation for each local image region by learning the internal distribution in an unsupervised manner via GAN. Instead of assuming degradation are spatially invariant across the whole image, we learn correction filters to adjust degradations to known degradations in a spatially variant way by a novel linearly-assembled pixel degradation-adaptive regression module (DARM). DARM is lightweight and easy to optimize on a dictionary of multiple pre-defined filter bases. Extensive experiments on synthetic and real-world datasets verify the effectiveness of our method both qualitatively and quantitatively. Code can be available at: https://github.com/edbca/DARSR.

Deep learning is commonly used to produce impressive results in reconstructing HDR images from LDR images. LDR stack-based methods are used for single-image HDR reconstruction, generating an HDR image from a deep learning generated LDR stack. However, current methods generate the LDR stack with predetermined exposure values (EVs), which may limit the quality of HDR reconstruction. To address this, we propose the continuous exposure value representation (CEVR) model, which uses an implicit function to generate LDR images with arbitrary EVs, including those unseen during training. Our flexible approach generates a continuous stack with more images containing diverse EVs, significantly improving HDR reconstruction. We use a cycle training strategy to supervise the model in generating continuous EV LDR images without corresponding ground truths. Our CEVR model outperforms existing methods, as demonstrated by experimental results.

Owing to the large distribution gap between the heterogeneous data in Visible-Infrared Person Re-identification (VI Re-ID), we point out that existing paradigms often suffer from the inter-modal semantic misalignment issue and thus fail to align and compare local details properly. In this paper, we present Concordant Attention Learning (CAL), a novel framework that learns semantic-aligned representations for VI Re-ID. Specifically, we design the Target-aware Concordant Alignment paradigm, which allows target-aware attention adaptation when aligning heterogeneous samples (i.e., adaptive attention adjustment according to the target image being aligned). This is achieved by exploiting the discriminative clues from the modality counterpart and designing effective modality-agnostic correspondence searching strategies. To ensure semantic concordance during the cross-modal retrieval stage, we further propose MatchDistill, which matches the attention patterns across modalities and learns their underlying semantic correlations by bipartite-graph-based similarity modeling and cross-modal knowledge exchange. Extensive experiments on VI Re-ID benchmark datasets demonstrate the effectiveness and superiority of the proposed CAL.

Recent advances in foundation models present new opportunities for interpretable visual recognition -- one can first query Large Language Models (LLMs) to obtain a set of attributes that describe each class, then apply vision-language models to classify images via these attributes. Pioneering work shows that querying thousands of attributes can achieve performance competitive with image features. However, our further investigation on 8 datasets reveals that LLM-generated attributes in a large quantity perform almost the same as random words. This surprising finding suggests that significant noise may be present in these attributes. We hypothesize that there exist subsets of attributes that can maintain the classification performance with much smaller sizes, and propose a novel learning-to-search method to discover those concise sets of attributes. As a result, on the CUB dataset, our method achieves performance close to that of massive LLM-generated attributes (e.g., 10k attributes for CUB), yet using only 32 attributes in total to distinguish 200 bird species. Furthermore, our new paradigm demonstrates several additional benefits: higher interpretability and interactivity for humans, and the ability to summarize knowledge for a recognition task.

Long-Term Person Re-Identification (LT-ReID) has become increasingly crucial in computer vision and biometrics. In this work, we aim to extend LT-ReID beyond pedestrian recognition to include a wider range of real-world human activities while still accounting for cloth-changing scenarios over large time gaps. This setting poses additional challenges due to the geometric misalignment and appearance ambiguity caused by the diversity of human pose and clothing. To address these challenges, we propose a new approach 3DInvarReID for (i) disentangling identity from non-identity components (pose, clothing shape, and texture) of 3D clothed humans, and (ii) reconstructing accurate 3D clothed body shapes and learning discriminative features of naked body shapes for person ReID in a joint manner. To better evaluate our study of LT-ReID, we collect a real-world dataset called CCDA, which contains a wide variety of human activities and clothing changes. Experimentally, we show the superior performance of our approach for person ReID.

Graph convolutional networks (GCNs) enable end-to-end learning on graph structured data. However, many works assume a given graph structure. When the input graph is noisy or unavailable, one approach is to construct or learn a latent graph structure. These methods typically fix the choice of node degree for the entire graph, which is suboptimal. Instead, we propose a novel end-to-end differentiable graph generator which builds graph topologies where each node selects both its neighborhood and its size. Our module can be readily integrated into existing pipelines involving graph convolution operations, replacing the predetermined or existing adjacency matrix with one that is learned, and optimized, as part of the general objective. As such it is applicable to any GCN. We integrate our module into trajectory prediction, point cloud classification and node classification pipelines resulting in improved accuracy over other structure-learning methods across a wide range of datasets and GCN backbones.

The success of deep neural networks for pan-sharpening is commonly in a form of black box, lacking transparency and interpretability. To alleviate this issue, we propose a novel model-driven deep unfolding framework with image reasoning prior tailored for the pan-sharpening task. Different from existing unfolding solutions that deliver the proximal operator networks as the uncertain and vague priors, our framework is motivated by the content reasoning ability of masked autoencoders (MAE) with insightful designs. Specifically, the pre-trained MAE with spatial masking strategy, acting as intrinsic reasoning prior, is embedded into unfolding architecture. Meanwhile, the pre-trained MAE with spatial-spectral masking strategy is treated as the regularization term within loss function to constrain the spatial-spectral consistency. Such designs penetrate the image reasoning prior into deep unfolding networks while improving its interpretability and representation capability. The uniqueness of our framework is that the holistic learning process is explicitly integrated with the inherent physical mechanism underlying the pan-sharpening task. Extensive experiments on multiple satellite datasets demonstrate the superiority of our method over the existing state-of-the-art approaches.

Single-photon 3D cameras can record the time-of-arrival of billions of photons per second with picosecond accuracy. One common approach to summarize the photon data stream is to build a per-pixel timestamp histogram, resulting in a 3D histogram tensor that encodes distances along the time axis. As the spatio-temporal resolution of the histogram tensor increases, the in-pixel memory requirements and output data rates can quickly become impractical. To overcome this limitation, we propose a family of linear compressive representations of histogram tensors that can be computed efficiently, in an online fashion, as a matrix operation. We design practical lightweight compressive representations that are amenable to an in-pixel implementation and consider the spatio-temporal information of each timestamp. Furthermore, we implement our proposed framework as the first layer of a neural network, which enables the joint end-to-end optimization of the compressive representations and a downstream SPAD data processing model. We find that a well-designed compressive representation can reduce in-sensor memory and data rates up to 2 orders of magnitude without significantly reducing 3D imaging quality. Finally, we analyze the power consumption implications through an on-chip implementation.

Predicting diverse yet admissible trajectories that adhere to the map constraints is challenging. Graph-based scene encoders have been proven effective for preserving local structures of maps by defining lane-level connections. However, such encoders do not capture more complex patterns emerging from long-range heterogeneous connections between nonadjacent interacting lanes. To this end, we shed new light on learning common driving patterns by introducing meTA ROad paTh (TAROT) to formulate combinations of various relations between lanes on the road topology. Intuitively, this can be viewed as finding feasible routes. Furthermore, we propose MEta-road NeTwORk (MENTOR) that helps trajectory prediction by providing it with TAROT as navigation tips. More specifically, 1) we define TAROT prediction as a novel self-supervised proxy task to identify the complex heterogeneous structure of the map. 2) For typical driving actions, we establish several TAROTs that result in multiple Heterogeneous Structure Learning (HSL) tasks. These tasks are used in MENTOR, which performs meta-learning by simultaneously predicting trajectories along with proxy tasks, identifying an optimal combination of them, and automatically balancing them to improve the primary task. We show that our model achieves state-of-the-art performance on the Argoverse dataset, especially on diversity and admissibility metrics, achieving up to 20% improvements in challenging scenarios. We further investigate the contribution of proposed modules in ablation studies.

The success of deep learning models has led to their adaptation and adoption by prominent video understanding methods. The majority of these approaches encode features in a joint space-time modality for which the inner workings and learned representations are difficult to visually interpret. We propose LEArned Preconscious Synthesis (LEAPS), an architecture-independent method for synthesizing videos from the internal spatiotemporal representations of models. Using a stimulus video and a target class, we prime a fixed space-time model and iteratively optimize a video initialized with random noise. Additional regularizers are used to improve the feature diversity of the synthesized videos alongside the cross-frame temporal coherence of motions. We quantitatively and qualitatively evaluate the applicability of LEAPS by inverting a range of spatiotemporal convolutional and attention-based architectures trained on Kinetics-400, which to the best of our knowledge has not been previously accomplished.

We focus on learning descriptive geometry and motion features from 4D point cloud sequences in this work. Existing works usually develop generic 4D learning tools without leveraging the prior that a 4D sequence comes from a single 3D scene with local dynamics. Based on this observation, we propose to learn region-wise coordinate frames that transform together with the underlying geometry. With such frames, we can factorize geometry and motion to facilitate a feature-space geometric reconstruction for more effective 4D learning. To learn such region frames, we develop a rotation equivariant network with a frame stabilization strategy. To leverage such frames for better spatial-temporal feature learning, we develop a frame-guided 4D learning scheme. Experiments show that this approach significantly outperforms previous state-of-the-art methods on a wide range of 4D understanding benchmarks.

Creating graphic layouts is a fundamental step in graphic designs. In this work, we present a novel generative model named LayoutDiffusion for automatic layout generation. As layout is typically represented as a sequence of discrete tokens, LayoutDiffusion models layout generation as a discrete denoising diffusion process. It learns to reverse a mild forward process, in which layouts become increasingly chaotic with the growth of forward steps and layouts in the neighboring steps do not differ too much. Designing such a mild forward process is however very challenging as layout has both categorical attributes and ordinal attributes. To tackle the challenge, we summarize three critical factors for achieving a mild forward process for the layout, i.e., legality, coordinate proximity and type disruption. Based on the factors, we propose a block-wise transition matrix coupled with a piece-wise linear noise schedule. Experiments on RICO and PubLayNet datasets show that LayoutDiffusion outperforms state-of-the-art approaches significantly. Moreover, it enables two conditional layout generation tasks in a plug-and-play manner without re-training and achieves better performance than existing methods. Project page: https://layoutdiffusion.github.io.

Most research on facial expression recognition (FER) is conducted in highly controlled environments, but its performance is often unacceptable when applied to real-world situations. This is because when unexpected objects occlude the face, the FER network faces difficulties extracting facial features and accurately predicting facial expressions. Therefore, occluded FER (OFER) is a challenging problem. Previous studies on occlusion-aware FER have typically required fully annotated facial images for training. However, collecting facial images with various occlusions and expression annotations is time-consuming and expensive. Latent-OFER, the proposed method, can detect occlusions, restore occluded parts of the face as if they were unoccluded, and recognize them, improving FER accuracy. This approach involves three steps: First, the vision transformer (ViT)-based occlusion patch detector masks the occluded position by training only latent vectors from the unoccluded patches using the support vector data description algorithm. Second, the hybrid reconstruction network generates the masking position as a complete image using the ViT and convolutional neural network (CNN). Last, the expression-relevant latent vector extractor retrieves and uses expression-related information from all latent vectors by applying a CNN-based class activation map. This mechanism has a significant advantage in preventing performance degradation from occlusion by unseen objects. The experimental results on several databases demonstrate the superiority of the proposed method over state-of-the-art methods.

Sample selection is a prevalent method in learning with noisy labels, where small-loss data are typically considered as correctly labeled data. However, this method may not effectively identify clean hard examples with large losses, which are critical for achieving the model's closeto-optimal generalization performance. In this paper, we propose a new framework, Late Stopping, which leverages the intrinsic robust learning ability of DNNs through a prolonged training process. Specifically, Late Stopping gradually shrinks the noisy dataset by removing high-probability mislabeled examples while retaining the majority of clean hard examples in the training set throughout the learning process. We empirically observe that mislabeled and clean examples exhibit differences in the number of epochs required for them to be consistently and correctly classified, and thus high-probability mislabeled examples can be removed. Experimental results on benchmark-simulated and real-world noisy datasets demonstrate that the proposed method outperforms state-of-the-art counterparts.

Location determination finds wide applications in daily life. Instead of existing efforts devoted to localizing tourist photos captured by perspective cameras, in this article, we focus on developing person positioning solutions using overhead fisheye cameras. Such solutions are advantageous in large field of view (FOV), low cost, anti-occlusion, and unaggressive work mode (without the necessity of cameras carried by persons). However, related studies are quite scarce, due to the paucity of data. To stimulate research in this exciting area, we present LOAF, the first large-scale overhead fisheye dataset for person detection and localization. LOAF is built with many essential features, e.g., i) the data cover abundant diversities in scenes, human pose, density, and location; ii) it contains currently the largest number of annotated pedestrian, i.e., 600K bounding boxes with ground-truth location information; iii) the body-boxes are labeled as radius-aligned so as to fully address the positioning challenge. To approach localization, we build a fisheye person detection network, which exploits the fisheye distortions by a clever position embedding strategy and is trained to predict radius-aligned human boxes end-to-end. Then, the actual locations of the detected persons are calculated by a numerical solution on the fisheye model and camera altitude data. Extensive experiments on LOAF validate the superiority of our fisheye detector w.r.t. previous methods, and show that our whole fisheye positioning solution is able to locate all persons in FOV with an accuracy of 0.5m, within 0.1s. Our dataset and code shall be released.

Semi-supervised learning has attracted increasing attention in the large-scale land cover mapping task. However, existing methods overlook the potential to alleviate the class imbalance problem by selecting a suitable set of unlabeled data. Besides, in class-imbalanced scenarios, existing pseudo-labeling methods mostly only pick confident samples, failing to exploit the hard samples during training. To tackle these issues, we propose a unified Class-Aware Semi-Supervised Semantic Segmentation framework. The proposed framework consists of three key components. To construct a better semi-supervised learning dataset, we propose a class-aware unlabeled data selection method that is more balanced towards the minority classes. Based on the built dataset with improved class balance, we propose a Class-Balanced Cross Entropy loss, jointly considering the annotation bias and the class bias to re-weight the loss in both sample and class levels to alleviate the class imbalance problem. Moreover, we propose the Class Center Contrast method to jointly utilize the labeled and unlabeled data. Specifically, we decompose the feature embedding space using the ground truth and pseudo-labels, and employ the embedding centers for hard and easy samples of each class per image in the contrast loss to exploit the hard samples during training. Compared with state-of-the-art class-balanced pseudo-labeling methods, the proposed method improves the mean accuracy and mIoU by 4.28% and 1.70%, respectively, on the large-scale Sentinel-2 dataset with 24 land cover classes.

Recent research on remote sensing object detection has largely focused on improving the representation of oriented bounding boxes but has overlooked the unique prior knowledge presented in remote sensing scenarios. Such prior knowledge can be useful because tiny remote sensing objects may be mistakenly detected without referencing a sufficiently long-range context, which can vary for different objects. This paper considers these priors and proposes the lightweight Large Selective Kernel Network (LSKNet). LSKNet can dynamically adjust its large spatial receptive field to better model the ranging context of various objects in remote sensing scenarios. To our knowledge, large and selective kernel mechanisms have not been previously explored in remote sensing object detection. Without bells and whistles, our lightweight LSKNet sets new state-of-the-art scores on standard benchmarks, i.e., HRSC2016 (98.46% mAP), DOTA-v1.0 (81.85% mAP), and FAIR1M-v1.0 (47.87% mAP).