Point cloud analysis has received much attention recently; and segmentation is one of the most important tasks. The success existing approaches attributed to deep network design large amount labelled training data, where latter assumed be always available. However, obtaining 3d point labels often very costly in practice. In this work, we propose a weakly supervised approach which requires only tiny fraction points stage. This made possible by learning gradient approximation exploitation...

The ability to understand the ways interact with objects from visual cues, a.k.a. affordance, is essential vision-guided robotic research. This involves categorizing, segmenting and reasoning of affordance. Relevant studies in 2D 2.5D image domains have been made previously, however, a truly functional understanding object affordance requires learning prediction 3D physical domain, which still absent community. In this work, we present AffordanceNet dataset, bench-mark 23k shapes 23 semantic...

State-of-the-art methods for semantic segmentation are based on deep neural networks that known to be data-hungry. Region-based active learning has shown a promising method reducing data annotation costs. A key design choice region-based AL is whether use regularly-shaped regions (e.g., rectangles) or irregularly-shaped region superpixels). In this work, we address question under realistic, click-based measurement of particular, revisit the super-pixels and demonstrate inappropriate cost...

Object detection in remote sensing images relies on a large amount of labeled data for training. However, the increasing number new categories and class imbalance make exhaustive annotation impractical. Few-shot object (FSOD) addresses this issue by leveraging meta-learning seen base classes fine-tuning novel with limited samples. Nonetheless, substantial scale orientation variations objects pose significant challenges to existing few-shot methods. To overcome these challenges, we propose...

Addressing the annotation challenge in 3D Point Cloud segmentation has inspired research into weakly supervised learning. Existing approaches mainly focus on exploiting manifold and pseudo-labeling to make use of large unlabeled data points. A fundamental here lies intra-class variations local geometric structure, resulting subclasses within a semantic class. In this work, we leverage intuition opt for maintaining an individual classifier each subclass. Technically, design multi-prototype...

Test-Time Adaptation aims to adapt source domain model testing data at inference stage with success demonstrated in adapting unseen corruptions. However, these attempts may fail under more challenging real-world scenarios. Existing works mainly consider test-time adaptation non-i.i.d. stream and continual shift. In this work, we first complement the existing TTA protocol a globally class imbalanced set. We demonstrate that combining all settings together poses new challenges methods. argue...

Many real-world video sequences cannot be conveniently categorized as general or degenerate; in such cases, imposing a false dichotomy using the fundamental matrix homography model for motion segmentation on would lead to difficulty. Even when we are confronted with scene-motion, approach still suffers from several defects, which discuss this paper. The full potential of could only realized if judiciously harness information simpler model. From these considerations, propose multi-model...

Deploying models on target domain data subject to distribution shift requires adaptation. Test-time training (TTT) emerges as a solution this adaptation under realistic scenario where access full source is not available, and instant inference the required. Despite many efforts into TTT, there confusion over experimental settings, thus leading unfair comparisons. In work, we first revisit TTT assumptions categorize protocols by two key factors, i.e. whether testing sequentially streamed model...

Existing approaches towards anomaly detection (AD) often rely on a substantial amount of anomaly-free data to train representation and density models. However, large datasets may not always be available before the inference stage; in which case an model must trained with only handful normal samples, a.k.a. few-shot (FSAD). In this paper, we propose novel methodology address challenge FSAD incorporates two important techniques. Firstly, employ pre-trained source dataset initialize weights....

Many real-world sequences cannot be conveniently categorized as general or degenerate; in such cases, imposing a false dichotomy using the fundamental matrix homography model for motion segmentation would lead to difficulty. Even when we are confronted with scene-motion, approach still suffers from several defects, which discuss this paper. The full potential of could only realized if judiciously harness information simpler model. From these considerations, propose multi-view spectral...

Recent work on curvilinear structure segmentation has mostly focused backbone network design and loss engineering. The challenge of collecting labelled data, an expensive labor intensive process, been overlooked. While data is to obtain, unlabelled often readily available. In this work, we propose SemiCurv, a semi-supervised learning (SSL) framework for that able utilize such reduce the labelling burden. Our addresses two key challenges in formulating manner. First, fully exploit power...

Point cloud analysis has received much attention recently; and segmentation is one of the most important tasks. The success existing approaches attributed to deep network design large amount labelled training data, where latter assumed be always available. However, obtaining 3d point labels often very costly in practice. In this work, we propose a weakly supervised approach which requires only tiny fraction points stage. This made possible by learning gradient approximation exploitation...

Deep learning models are the state-of-the-art methods for semantic point cloud segmentation, success of which relies on availability large-scale annotated datasets. However, it can be extremely time-consuming and prohibitively expensive to compile such In this work, we propose an active approach maximize model performance given limited annotation budgets. We investigate appropriate sample granularity selection under realistic cost measurement (clicks), demonstrate that super-point based...

Subspace clustering has been extensively studied from the hypothesis-and-test, algebraic, and spectral clustering-based perspectives. Most assume that only a single type/class of subspace is present. Generalizations to multiple types are non-trivial, plagued by challenges such as choice numbers models, sampling imbalance parameter tuning. In many real world problems, data may not lie perfectly on linear hand designed models fit into these situations. this work, we formulate multi-type...

3D object detection has recently received much attention due to its great potential in autonomous vehicle (AV). The success of deep learning based detectors relies on the availability large-scale annotated datasets, which is time-consuming and expensive compile, especially for bounding box annotation. In this work, we investigate diversity-based active (AL) as a solution alleviate annotation burden. Given limited budget, only most informative frames objects are automatically selected human...

Stereo-LiDAR fusion is often used for autonomous systems such as self-driving cars the two modalities are complementary to each other. Existing stereo-LiDAR methods mostly at feature level or outcome level, without considering uncertainty of depth estimation in modality. To this end, we propose a holistic and contextual evidential network (HCENet) estimation, which considers both intra-modality inter-modality uncertainties from stereo matching LiDAR point cloud completion. We design dual...

State-of-the-art methods for semantic segmentation are based on deep neural networks trained large-scale labeled datasets. Acquiring such datasets would incur large annotation costs, especially dense pixel-level prediction tasks like segmentation. We consider region-based active learning as a strategy to reduce costs while maintaining high performance. In this setting, batches of informative image regions instead entire images selected labeling. Importantly, we propose that enforcing local...

Background clutters pose challenges to defocus blur detection. Existing approaches often produce artifact predictions in background areas with clutter and relatively low confident boundary areas. In this work, we tackle the above issues from two perspectives. Firstly, inspired by recent success of self-attention mechanism, introduce channel-wise spatial-wise attention modules attentively aggregate features at different channels spatial locations obtain more discriminative features. Secondly,...