Tiny Object Detection with Single Point Supervision

Tiny objects, with their limited spatial resolution, often resemble point-like distributions. As a result, bounding box prediction using point-level supervision emerges as a natural and cost-effective alternative to traditional box-level supervision. However, the small scale and lack of distinctive features of tiny objects make point annotations prone to noise, posing significant hurdles for model robustness. To tackle these challenges, we propose Point Teacher--the first end-to-end point-supervised method for robust tiny object detection in aerial images. To handle label noise from scale ambiguity and location shifts in point annotations, Point Teacher employs the teacher-student architecture and decouples the learning into a two-phase denoising process. In this framework, the teacher network progressively denoises the pseudo boxes derived from noisy point annotations, guiding the student network's learning. Specifically, in the first phase, random masking of image regions facilitates regression learning, enabling the teacher to transform noisy point annotations into coarse pseudo boxes. In the second phase, these coarse pseudo boxes are refined using dynamic multiple instance learning, which adaptively selects the most reliable instance from dynamically constructed proposal bags around the coarse pseudo boxes. Extensive experiments on three tiny object datasets (i.e., AI-TOD-v2, SODA-A, and TinyPerson) validate the proposed method's effectiveness and robustness against point location shifts. Notably, relying solely on point supervision, our Point Teacher already shows comparable performance with box-supervised learning methods. Codes and models will be made publicly available.

QuadricsReg: Large-Scale Point Cloud Registration using Quadric Primitives

In the realm of large-scale point cloud registration, designing a compact symbolic representation is crucial for efficiently processing vast amounts of data, ensuring registration robustness against significant viewpoint variations and occlusions. This paper introduces a novel point cloud registration method, i.e., QuadricsReg, which leverages concise quadrics primitives to represent scenes and utilizes their geometric characteristics to establish correspondences for 6-DoF transformation estimation. As a symbolic feature, the quadric representation fully captures the primary geometric characteristics of scenes, which can efficiently handle the complexity of large-scale point clouds. The intrinsic characteristics of quadrics, such as types and scales, are employed to initialize correspondences. Then we build a multi-level compatibility graph set to find the correspondences using the maximum clique on the geometric consistency between quadrics. Finally, we estimate the 6-DoF transformation using the quadric correspondences, which is further optimized based on the quadric degeneracy-aware distance in a factor graph, ensuring high registration accuracy and robustness against degenerate structures. We test on 5 public datasets and the self-collected heterogeneous dataset across different LiDAR sensors and robot platforms. The exceptional registration success rates and minimal registration errors demonstrate the effectiveness of QuadricsReg in large-scale point cloud registration scenarios. Furthermore, the real-world registration testing on our self-collected heterogeneous dataset shows the robustness and generalization ability of QuadricsReg on different LiDAR sensors and robot platforms. The codes and demos will be released at \url{https://levenberg.github.io/QuadricsReg}.

MCVO: A Generic Visual Odometry for Arbitrarily Arranged Multi-Cameras

Making multi-camera visual SLAM systems easier to set up and more robust to the environment is always one of the focuses of vision robots. Existing monocular and binocular vision SLAM systems have narrow FoV and are fragile in textureless environments with degenerated accuracy and limited robustness. Thus multi-camera SLAM systems are gaining attention because they can provide redundancy for texture degeneration with wide FoV. However, current multi-camera SLAM systems face massive data processing pressure and elaborately designed camera configurations, leading to estimation failures for arbitrarily arranged multi-camera systems. To address these problems, we propose a generic visual odometry for arbitrarily arranged multi-cameras, which can achieve metric-scale state estimation with high flexibility in the cameras' arrangement. Specifically, we first design a learning-based feature extraction and tracking framework to shift the pressure of CPU processing of multiple video streams. Then we use the rigid constraints between cameras to estimate the metric scale poses for robust SLAM system initialization. Finally, we fuse the features of the multi-cameras in the SLAM back-end to achieve robust pose estimation and online scale optimization. Additionally, multi-camera features help improve the loop detection for pose graph optimization. Experiments on KITTI-360 and MultiCamData datasets validate the robustness of our method over arbitrarily placed cameras. Compared with other stereo and multi-camera visual SLAM systems, our method obtains higher pose estimation accuracy with better generalization ability. Our codes and online demos are available at \url{https://github.com/JunhaoWang615/MCVO}

LaVIDE: A Language-Vision Discriminator for Detecting Changes in Satellite Image with Map References

Change detection, which typically relies on the comparison of bi-temporal images, is significantly hindered when only a single image is available. Comparing a single image with an existing map, such as OpenStreetMap, which is continuously updated through crowd-sourcing, offers a viable solution to this challenge. Unlike images that carry low-level visual details of ground objects, maps convey high-level categorical information. This discrepancy in abstraction levels complicates the alignment and comparison of the two data types. In this paper, we propose a \textbf{La}nguage-\textbf{VI}sion \textbf{D}iscriminator for d\textbf{E}tecting changes in satellite image with map references, namely \ours{}, which leverages language to bridge the information gap between maps and images. Specifically, \ours{} formulates change detection as the problem of ``{\textit Does the pixel belong to [class]?}'', aligning maps and images within the feature space of the language-vision model to associate high-level map categories with low-level image details. Moreover, we build a mixture-of-experts discriminative module, which compares linguistic features from maps with visual features from images across various semantic perspectives, achieving comprehensive semantic comparison for change detection. Extensive evaluation on four benchmark datasets demonstrates that \ours{} can effectively detect changes in satellite image with map references, outperforming state-of-the-art change detection algorithms, e.g., with gains of about $13.8$\% on the DynamicEarthNet dataset and $4.3$\% on the SECOND dataset.

Intensity Field Decomposition for Tissue-Guided Neural Tomography

Cone-beam computed tomography (CBCT) typically requires hundreds of X-ray projections, which raises concerns about radiation exposure. While sparse-view reconstruction reduces the exposure by using fewer projections, it struggles to achieve satisfactory image quality. To address this challenge, this article introduces a novel sparse-view CBCT reconstruction method, which empowers the neural field with human tissue regularization. Our approach, termed tissue-guided neural tomography (TNT), is motivated by the distinct intensity differences between bone and soft tissue in CBCT. Intuitively, separating these components may aid the learning process of the neural field. More precisely, TNT comprises a heterogeneous quadruple network and the corresponding training strategy. The network represents the intensity field as a combination of soft and hard tissue components, along with their respective textures. We train the network with guidance from estimated tissue projections, enabling efficient learning of the desired patterns for the network heads. Extensive experiments demonstrate that the proposed method significantly improves the sparse-view CBCT reconstruction with a limited number of projections ranging from 10 to 60. Our method achieves comparable reconstruction quality with fewer projections and faster convergence compared to state-of-the-art neural rendering based methods.

TAS: Distilling Arbitrary Teacher and Student via a Hybrid Assistant

Most knowledge distillation (KD) methodologies predominantly focus on teacher-student pairs with similar architectures, such as both being convolutional neural networks (CNNs). However, the potential and flexibility of KD can be greatly improved by expanding it to novel Cross-Architecture KD (CAKD), where the knowledge of homogeneous and heterogeneous teachers can be transferred flexibly to a given student. The primary challenge in CAKD lies in the substantial feature gaps between heterogeneous models, originating from the distinction of their inherent inductive biases and module functions. To this end, we introduce an assistant model as a bridge to facilitate smooth feature knowledge transfer between heterogeneous teachers and students. More importantly, within our proposed design principle, the assistant model combines the advantages of cross-architecture inductive biases and module functions by merging convolution and attention modules derived from both student and teacher module functions. Furthermore, we observe that heterogeneous features exhibit diverse spatial distributions in CAKD, hindering the effectiveness of conventional pixel-wise mean squared error (MSE) loss. Therefore, we leverage a spatial-agnostic InfoNCE loss to align features after spatial smoothing, thereby improving the feature alignments in CAKD. Our proposed method is evaluated across some homogeneous model pairs and arbitrary heterogeneous combinations of CNNs, ViTs, and MLPs, achieving state-of-the-art performance for distilled models with a maximum gain of 11.47% on CIFAR-100 and 3.67% on ImageNet-1K. Our code and models will be released.

Partial Distribution Matching via Partial Wasserstein Adversarial Networks

This paper studies the problem of distribution matching (DM), which is a fundamental machine learning problem seeking to robustly align two probability distributions. Our approach is established on a relaxed formulation, called partial distribution matching (PDM), which seeks to match a fraction of the distributions instead of matching them completely. We theoretically derive the Kantorovich-Rubinstein duality for the partial Wasserstain-1 (PW) discrepancy, and develop a partial Wasserstein adversarial network (PWAN) that efficiently approximates the PW discrepancy based on this dual form. Partial matching can then be achieved by optimizing the network using gradient descent. Two practical tasks, point set registration and partial domain adaptation are investigated, where the goals are to partially match distributions in 3D space and high-dimensional feature space respectively. The experiment results confirm that the proposed PWAN effectively produces highly robust matching results, performing better or on par with the state-of-the-art methods.

Exploring Scene Coherence for Semi-Supervised 3D Semantic Segmentation

Semi-supervised semantic segmentation, which efficiently addresses the limitation of acquiring dense annotations, is essential for 3D scene understanding. Most methods leverage the teacher model to generate pseudo labels, and then guide the learning of the student model on unlabeled scenes. However, they focus only on points with pseudo labels while directly overlooking points without pseudo labels, namely intra-scene inconsistency, leading to semantic ambiguity. Moreover, inter-scene correlation between labeled and unlabeled scenes contribute to transferring rich annotation information, yet this has not been explored for the semi-supervised tasks. To address these two problems, we propose to explore scene coherence for semi-supervised 3D semantic segmentation, dubbed CoScene. Inspired by the unstructured and unordered nature of the point clouds, our CoScene adopts the straightforward point erasure strategy to ensure the intra-scene consistency. Moreover, patch-based data augmentation is proposed to enhance the inter-scene information transfer between labeled and unlabeled scenes at both scene and instance levels. Extensive experimental results on SemanticKITTI and nuScenes show that our approach outperforms existing methods.

DMTG: One-Shot Differentiable Multi-Task Grouping

We aim to address Multi-Task Learning (MTL) with a large number of tasks by Multi-Task Grouping (MTG). Given N tasks, we propose to simultaneously identify the best task groups from 2^N candidates and train the model weights simultaneously in one-shot, with the high-order task-affinity fully exploited. This is distinct from the pioneering methods which sequentially identify the groups and train the model weights, where the group identification often relies on heuristics. As a result, our method not only improves the training efficiency, but also mitigates the objective bias introduced by the sequential procedures that potentially lead to a suboptimal solution. Specifically, we formulate MTG as a fully differentiable pruning problem on an adaptive network architecture determined by an underlying Categorical distribution. To categorize N tasks into K groups (represented by K encoder branches), we initially set up KN task heads, where each branch connects to all N task heads to exploit the high-order task-affinity. Then, we gradually prune the KN heads down to N by learning a relaxed differentiable Categorical distribution, ensuring that each task is exclusively and uniquely categorized into only one branch. Extensive experiments on CelebA and Taskonomy datasets with detailed ablations show the promising performance and efficiency of our method. The codes are available at https://github.com/ethanygao/DMTG.

Towards Human-Level 3D Relative Pose Estimation: Generalizable, Training-Free, with Single Reference

Humans can easily deduce the relative pose of an unseen object, without label/training, given only a single query-reference image pair. This is arguably achieved by incorporating (i) 3D/2.5D shape perception from a single image, (ii) render-and-compare simulation, and (iii) rich semantic cue awareness to furnish (coarse) reference-query correspondence. Existing methods implement (i) by a 3D CAD model or well-calibrated multiple images and (ii) by training a network on specific objects, which necessitate laborious ground-truth labeling and tedious training, potentially leading to challenges in generalization. Moreover, (iii) was less exploited in the paradigm of (ii), despite that the coarse correspondence from (iii) enhances the compare process by filtering out non-overlapped parts under substantial pose differences/occlusions. Motivated by this, we propose a novel 3D generalizable relative pose estimation method by elaborating (i) with a 2.5D shape from an RGB-D reference, (ii) with an off-the-shelf differentiable renderer, and (iii) with semantic cues from a pretrained model like DINOv2. Specifically, our differentiable renderer takes the 2.5D rotatable mesh textured by the RGB and the semantic maps (obtained by DINOv2 from the RGB input), then renders new RGB and semantic maps (with back-surface culling) under a novel rotated view. The refinement loss comes from comparing the rendered RGB and semantic maps with the query ones, back-propagating the gradients through the differentiable renderer to refine the 3D relative pose. As a result, our method can be readily applied to unseen objects, given only a single RGB-D reference, without label/training. Extensive experiments on LineMOD, LM-O, and YCB-V show that our training-free method significantly outperforms the SOTA supervised methods, especially under the rigorous Acc@5/10/15{\deg} metrics and the challenging cross-dataset settings.