Deep learning has largely reshaped remote sensing research for aerial image understanding. Nevertheless, most of existing deep models are initialized with ImageNet pretrained weights, where the natural images inevitably presents a large domain gap relative to the aerial images, probably limiting the finetuning performance on downstream aerial scene tasks. This issue motivates us to conduct an empirical study of remote sensing pretraining (RSP). To this end, we train different networks from scratch with the help of the largest remote sensing scene recognition dataset up to now-MillionAID, to obtain the remote sensing pretrained backbones, including both convolutional neural networks (CNN) and vision transformers such as Swin and ViTAE, which have shown promising performance on computer vision tasks. Then, we investigate the impact of ImageNet pretraining (IMP) and RSP on a series of downstream tasks including scene recognition, semantic segmentation, object detection, and change detection using the CNN and vision transformers backbones. We have some empirical findings as follows. First, vision transformers generally outperforms CNN backbones, where ViTAE achieves the best performance, owing to its strong representation capacity by introducing intrinsic inductive bias from convolutions to transformers. Second, both IMP and RSP help deliver better performance, where IMP enjoys a versatility by learning more universal representations from diverse images belonging to much more categories while RSP is distinctive in perceiving remote sensing related semantics. Third, RSP mitigates the data discrepancy of IMP for remote sensing but may still suffer from the task discrepancy, where downstream tasks require different representations from the scene recognition task. These findings call for further research efforts on both large-scale pretraining datasets and effective pretraining methods.
This paper addresses the problem of document image dewarping, which aims at eliminating the geometric distortion in document images for document digitization. Instead of designing a better neural network to approximate the optical flow fields between the inputs and outputs, we pursue the best readability by taking the text lines and the document boundaries into account from a constrained optimization perspective. Specifically, our proposed method first learns the boundary points and the pixels in the text lines and then follows the most simple observation that the boundaries and text lines in both horizontal and vertical directions should be kept after dewarping to introduce a novel grid regularization scheme. To obtain the final forward mapping for dewarping, we solve an optimization problem with our proposed grid regularization. The experiments comprehensively demonstrate that our proposed approach outperforms the prior arts by large margins in terms of readability (with the metrics of Character Errors Rate and the Edit Distance) while maintaining the best image quality on the publicly-available DocUNet benchmark.
Modern object detectors have achieved impressive progress under the close-set setup. However, open-set object detection (OSOD) remains challenging since objects of unknown categories are often misclassified to existing known classes. In this work, we propose to identify unknown objects by separating high/low-density regions in the latent space, based on the consensus that unknown objects are usually distributed in low-density latent regions. As traditional threshold-based methods only maintain limited low-density regions, which cannot cover all unknown objects, we present a novel Open-set Detector (OpenDet) with expanded low-density regions. To this aim, we equip OpenDet with two learners, Contrastive Feature Learner (CFL) and Unknown Probability Learner (UPL). CFL performs instance-level contrastive learning to encourage compact features of known classes, leaving more low-density regions for unknown classes; UPL optimizes unknown probability based on the uncertainty of predictions, which further divides more low-density regions around the cluster of known classes. Thus, unknown objects in low-density regions can be easily identified with the learned unknown probability. Extensive experiments demonstrate that our method can significantly improve the OSOD performance, e.g., OpenDet reduces the Absolute Open-Set Errors by 25%-35% on six OSOD benchmarks. Code is available at: https://github.com/csuhan/opendet2.
Given two point sets, the problem of registration is to recover a transformation that matches one set to the other. This task is challenging due to the presence of the large number of outliers, the unknown non-rigid deformations and the large sizes of point sets. To obtain strong robustness against outliers, we formulate the registration problem as a partial distribution matching (PDM) problem, where the goal is to partially match the distributions represented by point sets in a metric space. To handle large point sets, we propose a scalable PDM algorithm by utilizing the efficient partial Wasserstein-1 (PW) discrepancy. Specifically, we derive the Kantorovich-Rubinstein duality for the PW discrepancy, and show its gradient can be explicitly computed. Based on these results, we propose a partial Wasserstein adversarial network (PWAN), which is able to approximate the PW discrepancy by a neural network, and minimize it by gradient descent. In addition, it also incorporates an efficient coherence regularizer for non-rigid transformations to avoid unrealistic deformations. We evaluate PWAN on practical point set registration tasks, and show that the proposed PWAN is robust, scalable and performs more favorably than the state-of-the-art methods.
Given an aerial image, aerial scene parsing (ASP) targets to interpret the semantic structure of the image content, e.g., by assigning a semantic label to every pixel of the image. With the popularization of data-driven methods, the past decades have witnessed promising progress on ASP by approaching the problem with the schemes of tile-level scene classification or segmentation-based image analysis, when using high-resolution aerial images. However, the former scheme often produces results with tile-wise boundaries, while the latter one needs to handle the complex modeling process from pixels to semantics, which often requires large-scale and well-annotated image samples with pixel-wise semantic labels. In this paper, we address these issues in ASP, with perspectives from tile-level scene classification to pixel-wise semantic labeling. Specifically, we first revisit aerial image interpretation by a literature review. We then present a large-scale scene classification dataset that contains one million aerial images termed Million-AID. With the presented dataset, we also report benchmarking experiments using classical convolutional neural networks (CNNs). Finally, we perform ASP by unifying the tile-level scene classification and object-based image analysis to achieve pixel-wise semantic labeling. Intensive experiments show that Million-AID is a challenging yet useful dataset, which can serve as a benchmark for evaluating newly developed algorithms. When transferring knowledge from Million-AID, fine-tuning CNN models pretrained on Million-AID perform consistently better than those pretrained ImageNet for aerial scene classification. Moreover, our designed hierarchical multi-task learning method achieves the state-of-the-art pixel-wise classification on the challenging GID, bridging the tile-level scene classification toward pixel-wise semantic labeling for aerial image interpretation.
Zero-shot semantic segmentation (ZS3) aims to segment the novel categories that have not been seen in the training. Existing works formulate ZS3 as a pixel-level zero-shot classification problem, and transfer semantic knowledge from seen classes to unseen ones with the help of language models pre-trained only with texts. While simple, the pixel-level ZS3 formulation shows the limited capability to integrate vision-language models that are often pre-trained with image-text pairs and currently demonstrate great potential for vision tasks. Inspired by the observation that humans often perform segment-level semantic labeling, we propose to decouple the ZS3 into two sub-tasks: 1) a class-agnostic grouping task to group the pixels into segments. 2) a zero-shot classification task on segments. The former sub-task does not involve category information and can be directly transferred to group pixels for unseen classes. The latter subtask performs at segment-level and provides a natural way to leverage large-scale vision-language models pre-trained with image-text pairs (e.g. CLIP) for ZS3. Based on the decoupling formulation, we propose a simple and effective zero-shot semantic segmentation model, called ZegFormer, which outperforms the previous methods on ZS3 standard benchmarks by large margins, e.g., 35 points on the PASCAL VOC and 3 points on the COCO-Stuff in terms of mIoU for unseen classes. Code will be released at https://github.com/dingjiansw101/ZegFormer.
Targeting at depicting land covers with pixel-wise semantic categories, semantic segmentation in remote sensing images needs to portray diverse distributions over vast geographical locations, which is difficult to be achieved by the homogeneous pixel-wise forward paths in the architectures of existing deep models. Although several algorithms have been designed to select pixel-wise adaptive forward paths for natural image analysis, it still lacks theoretical supports on how to obtain optimal selections. In this paper, we provide mathematical analyses in terms of the parameter optimization, which guides us to design a method called Hidden Path Selection Network (HPS-Net). With the help of hidden variables derived from an extra mini-branch, HPS-Net is able to tackle the inherent problem about inaccessible global optimums by adjusting the direct relationships between feature maps and pixel-wise path selections in existing algorithms, which we call hidden path selection. For the better training and evaluation, we further refine and expand the 5-class Gaofen Image Dataset (GID-5) to a new one with 15 land-cover categories, i.e., GID-15. The experimental results on both GID-5 and GID-15 demonstrate that the proposed modules can stably improve the performance of different deep structures, which validates the proposed mathematical analyses.
In this paper, we propose an end-to-end learning framework for event-based motion deblurring in a self-supervised manner, where real-world events are exploited to alleviate the performance degradation caused by data inconsistency. To achieve this end, optical flows are predicted from events, with which the blurry consistency and photometric consistency are exploited to enable self-supervision on the deblurring network with real-world data. Furthermore, a piece-wise linear motion model is proposed to take into account motion non-linearities and thus leads to an accurate model for the physical formation of motion blurs in the real-world scenario. Extensive evaluation on both synthetic and real motion blur datasets demonstrates that the proposed algorithm bridges the gap between simulated and real-world motion blurs and shows remarkable performance for event-based motion deblurring in real-world scenarios.
This paper presents a method of learning Local-GlObal Contextual Adaptation for fully end-to-end and fast bottom-up human Pose estimation, dubbed as LOGO-CAP. It is built on the conceptually simple center-offset formulation that lacks inaccuracy for pose estimation. When revisiting the bottom-up human pose estimation with the thought of "thinking, fast and slow" by D. Kahneman, we introduce a "slow keypointer" to remedy the lack of sufficient accuracy of the "fast keypointer". In learning the "slow keypointer", the proposed LOGO-CAP lifts the initial "fast" keypoints by offset predictions to keypoint expansion maps (KEMs) to counter their uncertainty in two modules. Firstly, the local KEMs (e.g., 11x11) are extracted from a low-dimensional feature map. A proposed convolutional message passing module learns to "re-focus" the local KEMs to the keypoint attraction maps (KAMs) by accounting for the structured output prediction nature of human pose estimation, which is directly supervised by the object keypoint similarity (OKS) loss in training. Secondly, the global KEMs are extracted, with a sufficiently large region-of-interest (e.g., 97x97), from the keypoint heatmaps that are computed by a direct map-to-map regression. Then, a local-global contextual adaptation module is proposed to convolve the global KEMs using the learned KAMs as the kernels. This convolution can be understood as the learnable offsets guided deformable and dynamic convolution in a pose-sensitive way. The proposed method is end-to-end trainable with near real-time inference speed, obtaining state-of-the-art performance on the COCO keypoint benchmark for bottom-up human pose estimation. With the COCO trained model, our LOGO-CAP also outperforms prior arts by a large margin on the challenging OCHuman dataset.