LiDAR segmentation is crucial for autonomous driving perception. Recent trends favor point- or voxel-based methods as they often yield better performance than the traditional range view representation. In this work, we unveil several key factors in building powerful range view models. We observe that the "many-to-one" mapping, semantic incoherence, and shape deformation are possible impediments against effective learning from range view projections. We present RangeFormer -- a full-cycle framework comprising novel designs across network architecture, data augmentation, and post-processing -- that better handles the learning and processing of LiDAR point clouds from the range view. We further introduce a Scalable Training from Range view (STR) strategy that trains on arbitrary low-resolution 2D range images, while still maintaining satisfactory 3D segmentation accuracy. We show that, for the first time, a range view method is able to surpass the point, voxel, and multi-view fusion counterparts in the competing LiDAR semantic and panoptic segmentation benchmarks, i.e., SemanticKITTI, nuScenes, and ScribbleKITTI.
论文检索
输入标题、作者或关键词，从 2,156 篇学术成果中精准定位
Learning per-point semantic features from the hierarchical feature pyramid is essential for point cloud semantic segmentation. However, most previous methods suffered from ambiguous region features or failed to refine per-point features effectively, which leads to information loss and ambiguous semantic identification. To resolve this, we propose Retro-FPN to model the per-point feature prediction as an explicit and retrospective refining process, which goes through all the pyramid layers to extract semantic features explicitly for each point. Its key novelty is a retro-transformer for summarizing semantic contexts from the previous layer and accordingly refining the features in the current stage. In this way, the categorization of each point is conditioned on its local semantic pattern. Specifically, the retro-transformer consists of a local cross-attention block and a semantic gate unit. The cross-attention serves to summarize the semantic pattern retrospectively from the previous layer. And the gate unit carefully incorporates the summarized contexts and refines the current semantic features. Retro-FPN is a pluggable neural network that applies to hierarchical decoders. By integrating Retro-FPN with three representative backbones, including both point-based and voxel-based methods, we show that Retro-FPN can significantly improve performance over state-of-the-art backbones. Comprehensive experiments on widely used benchmarks can justify the effectiveness of our design. The source is available at https://github.com/AllenXiangX/Retro-FPN.
S-TREK: Sequential Translation and Rotation Equivariant Keypoints for Local Feature Extraction
PDF ↗In this work we introduce S-TREK, a novel local feature extractor that combines a deep keypoint detector, which is both translation and rotation equivariant by design, with a lightweight deep descriptor extractor. We train the S-TREK keypoint detector within a framework inspired by reinforcement learning, where we leverage a sequential procedure to maximize a reward directly related to keypoint repeatability. Our descriptor network is trained following a "detect, then describe" approach, where the descriptor loss is evaluated only at those locations where keypoints have been selected by the already trained detector. Extensive experiments on multiple benchmarks confirm the effectiveness of our proposed method, with S-TREK often outperforming other state-of-the-art methods in terms of repeatability and quality of the recovered poses, especially when dealing with in-plane rotations.
This paper introduces the first two landmark pixel retrieval benchmarks. Like semantic segmentation extends classification to the pixel level, pixel retrieval is an extension of image retrieval and offers information about which pixels are related to the query object. In addition to retrieving images for the given query, it helps users quickly identify the query object in true positive images and exclude false positive images by denoting the correlated pixels. Our user study results show pixel-level annotation can significantly improve the user experience. Compared with semantic and instance segmentation, pixel retrieval requires a fine-grained recognition capability for variable-granularity targets. To this end, we propose pixel retrieval benchmarks named PROxford and PRParis, which are based on the widely used image retrieval datasets, ROxford and RParis. Three professional annotators label 5,942 images with two rounds of double-checking and refinement. Furthermore, we conduct extensive experiments and analysis on the SOTA methods in image search, image matching, detection, segmentation, and dense matching using our pixel retrieval benchmarks. Results show that the pixel retrieval task is challenging to these approaches and distinctive from existing problems, suggesting that further research can advance the content-based pixel-retrieval and, thus, user search experience.
Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive segmentation. To fill this gap, we propose SimpleClick, the first plain-backbone method for interactive segmentation. Other than the plain backbone, we also explore several variants of simple feature pyramid networks that only take as input the last feature representation of the backbone. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving 21.8% over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We further develop an extremely tiny ViT backbone for SimpleClick and provide a detailed computational analysis, highlighting its suitability as a practical annotation tool.
Although existing stereo matching models have achieved continuous improvement, they often face issues related to trustworthiness due to the absence of uncertainty estimation. Additionally, effectively leveraging multi-scale and multi-view knowledge of stereo pairs remains unexplored. In this paper, we introduce the Evidential Local-global Fusion (ELF) framework for stereo matching, which endows both uncertainty estimation and confidence-aware fusion with trustworthy heads. Instead of predicting the disparity map alone, our model estimates an evidential-based disparity considering both aleatoric and epistemic uncertainties. With the normal inverse-Gamma distribution as a bridge, the proposed framework realizes intra evidential fusion of multi-level predictions and inter evidential fusion between cost-volume-based and transformer-based stereo matching. Extensive experimental results show that the proposed framework exploits multi-view information effectively and achieves state-of-the-art overall performance both on accuracy and cross-domain generalization. The codes are available at https://github.com/jimmy19991222/ELFNet.
Semi-Supervised Learning (SSL) under class distribution mismatch aims to tackle a challenging problem wherein unlabeled data contain lots of unknown categories unseen in the labeled ones. In such mismatch scenarios, traditional SSL suffers severe performance damage due to the harmful invasion of the instances with unknown categories into the target classifier. In this study, by strict mathematical reasoning, we reveal that the SSL error under class distribution mismatch is composed of pseudo-labeling error and invasion error, both of which jointly bound the SSL population risk. To alleviate the SSL error, we propose a robust SSL framework called Weight-Aware Distillation (WAD) that, by weights, selectively transfers knowledge beneficial to the target task from unsupervised contrastive representation to the target classifier. Specifically, WAD captures adaptive weights and high-quality pseudo-labels to target instances by exploring point mutual information (PMI) in representation space to maximize the role of unlabeled data and filter unknown categories. Theoretically, we prove that WAD has a tight upper bound of population risk under class distribution mismatch. Experimentally, extensive results demonstrate that WAD outperforms five state-of-the-art SSL approaches and one standard baseline on two benchmark datasets, CIFAR10 and CIFAR100, and an artificial cross-dataset. The code is available at https://github.com/RUC-DWBI-ML/research/tree/main/WAD-master.
Modern deepfake detectors have achieved encouraging results, when training and test images are drawn from the same data collection. However, when these detectors are applied to images produced with unknown deepfake-generation techniques, considerable performance degradations are commonly observed. In this paper, we propose a novel deepfake detector, called SeeABLE, that formalizes the detection problem as a (one-class) out-of-distribution detection task and generalizes better to unseen deepfakes. Specifically, SeeABLE first generates local image perturbations (referred to as soft-discrepancies) and then pushes the perturbed faces towards predefined prototypes using a novel regression-based bounded contrastive loss. To strengthen the generalization performance of SeeABLE to unknown deepfake types, we generate a rich set of soft discrepancies and train the detector: (i) to localize, which part of the face was modified, and (ii) to identify the alteration type. To demonstrate the capabilities of SeeABLE, we perform rigorous experiments on several widely-used deepfake datasets and show that our model convincingly outperforms competing state-of-the-art detectors, while exhibiting highly encouraging generalization capabilities. The source code for SeeABLE is available from: https://github.com/anonymous-author-sub/seeable.
Label noise is one of the key factors that lead to the poor generalization of deep learning models. Existing label-noise learning methods usually assume that the ground-truth classes of the training data are balanced. However, the real-world data is often imbalanced, leading to the inconsistency between observed and intrinsic class distribution with label noises. In this case, it is hard to distinguish clean samples from noisy samples on the intrinsic tail classes with the unknown intrinsic class distribution. In this paper, we propose a learning framework for label-noise learning with intrinsically long-tailed data. Specifically, we propose two-stage bi-dimensional sample selection (TABASCO) to better separate clean samples from noisy samples, especially for the tail classes. TABASCO consists of two new separation metrics that complement each other to compensate for the limitation of using a single metric in sample separation. Extensive experiments on benchmarks demonstrate the effectiveness of our method. Our code is available at https://github.com/Wakings/TABASCO.
Implicit neural representations have shown powerful capacity in modeling real-world 3D scenes, offering superior performance in novel view synthesis. In this paper, we target a more challenging scenario, i.e., joint scene novel view synthesis and editing based on implicit neural scene representations. State-of-the-art methods in this direction typically consider building separate networks for these two tasks (i.e., view synthesis and editing). Thus, the modeling of interactions and correlations between these two tasks is very limited, which, however, is critical for learning high-quality scene representations.To tackle this problem, in this paper, we propose a unified Neural Radiance Field (NeRF) framework to effectively perform joint scene decomposition and composition for modeling real-world scenes. The decomposition aims at learning disentangled 3D representations of different objects and the background, allowing for scene editing, while scene composition models an entire scene representation for novel view synthesis. Specifically, with a two-stage NeRF framework, we learn a coarse stage for predicting a global radiance field as guidance for point sampling, and in the second fine-grained stage, we perform scene decomposition by a novel one-hot object radiance field regularization module and a pseudo supervision via inpainting to handle ambiguous background regions occluded by objects. The decomposed object-level radiance fields are further composed by using activations from the decomposition module. Extensive quantitative and qualitative results show the effectiveness of our method for scene decomposition and composition, outperforming state-of-the-art methods for both novel-view synthesis and editing tasks.
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.
FCCNs: Fully Complex-valued Convolutional Networks using Complex-valued Color Model and Loss Function
PDF ↗Although complex-valued convolutional neural networks (iCNNs) have existed for a while, they lack proper complex-valued image inputs and loss functions. In addition, all their operations are not complex-valued as they have both complex-valued convolutional layers and real-valued fully-connected layers. As a result, they lack an end-to-end flow of complex-valued information, making them inconsistent w.r.t. the claimed operating domain, i.e., complex numbers. Considering these inconsistencies, we propose a complex-valued color model and loss function and turn fully-connected layers into convolutional layers. All these contributions culminate in what we call FCCNs (Fully Complex-valued Convolutional Networks), which take complex-valued images as inputs, perform only complex-valued operations, and have a complex-valued loss function. Thus, our proposed FCCNs have an end-to-end flow of complex-valued information, which lacks in existing iCNNs. Our extensive experiments on five image classification benchmark datasets show that FCCNs consistently perform better than existing iCNNs. Code is available at https://github.com/saurabhya/FCCNs .
Deep Neural Networks (DNNs)-based semantic segmentation models trained on a source domain often struggle to generalize to unseen target domains, i.e., a domain gap problem. Texture often contributes to the domain gap, making DNNs vulnerable to domain shift because they are prone to be texture-biased. Existing Domain Generalized Semantic Segmentation (DGSS) methods have alleviated the domain gap problem by guiding models to prioritize shape over texture. On the other hand, shape and texture are two prominent and complementary cues in semantic segmentation. This paper argues that leveraging texture is crucial for improving performance in DGSS. Specifically, we propose a novel framework, coined Texture Learning Domain Randomization (TLDR). TLDR includes two novel losses to effectively enhance texture learning in DGSS: (1) a texture regularization loss to prevent overfitting to source domain textures by using texture features from an ImageNet pre-trained model and (2) a texture generalization loss that utilizes random style images to learn diverse texture representations in a self-supervised manner. Extensive experimental results demonstrate the superiority of the proposed TLDR; e.g., TLDR achieves 46.5 mIoU on GTA-to-Cityscapes using ResNet-50, which improves the prior state-of-the-art method by 1.9 mIoU. The source code is available at https://github.com/ssssshwan/TLDR.
With frame-based cameras, capturing fast-moving scenes without suffering from blur often comes at the cost of low SNR and low contrast. Worse still, the photometric constancy that enhancement techniques heavily relied on is fragile for frames with short exposure. Event cameras can record brightness changes at an extremely high temporal resolution. For low-light videos, event data are not only suitable to help capture temporal correspondences but also provide alternative observations in the form of intensity ratios between consecutive frames and exposure-invariant information. Motivated by this, we propose a low-light video enhancement method with hybrid inputs of events and frames. Specifically, a neural network is trained to establish spatiotemporal coherence between visual signals with different modalities and resolutions by constructing correlation volume across space and time. Experimental results on synthetic and real data demonstrate the superiority of the proposed method compared to the state-of-the-art methods.
Video frame interpolation (VFI) is a very active research topic due to its broad applicability to many applications, including video enhancement, video encoding, and slow-motion effects. VFI methods have been advanced by improving the overall image quality for challenging sequences containing occlusions, large motion, and dynamic texture. This mainstream research direction neglects that foreground and background regions have different importance in perceptual image quality. Moreover, accurate synthesis of moving objects can be of utmost importance in computer vision applications. In this paper, we propose a video object segmentation (VOS)-aware training framework called VOS-VFI that allows VFI models to interpolate frames with more precise object boundaries. Specifically, we exploit VOS as an auxiliary task to help train VFI models by providing additional loss functions, including segmentation loss and bi-directional consistency loss. From extensive experiments, we demonstrate that VOS-VFI can boost the performance of existing VFI models by rendering clear object boundaries. Moreover, VOS-VFI displays its effectiveness on multiple benchmarks for different applications, including video object segmentation, object pose estimation, and visual tracking.
The Devil is in the Upsampling: Architectural Decisions Made Simpler for Denoising with Deep Image Prior
PDF ↗Deep Image Prior (DIP) shows that some network architectures inherently tend towards generating smooth images while resisting noise, a phenomenon known as spectral bias. Image denoising is a natural application of this property. Although denoising with DIP mitigates the need for large training sets, two often intertwined practical challenges need to be overcome: architectural design and noise fitting. Existing methods either handcraft or search for suitable architectures from a vast design space, due to the limited understanding of how architectural choices affect the denoising outcome. In this study, we demonstrate from a frequency perspective that unlearnt upsampling is the main driving force behind the denoising phenomenon with DIP. This finding leads to straightforward strategies for identifying a suitable architecture for every image without laborious search. Extensive experiments show that the estimated architectures achieve superior denoising results than existing methods with up to 95% fewer parameters. Thanks to this under-parameterization, the resulting architectures are less prone to noise-fitting.
Current high-performance semantic segmentation models are purely data-driven sub-symbolic approaches and blind to the structured nature of the visual world. This is in stark contrast to human cognition which abstracts visual perceptions at multiple levels and conducts symbolic reasoning with such structured abstraction. To fill these fundamental gaps, we devise LogicSeg, a holistic visual semantic parser that integrates neural inductive learning and logic reasoning with both rich data and symbolic knowledge. In particular, the semantic concepts of interest are structured as a hierarchy, from which a comprehensive set of constraints are derived for describing the symbolic relations and formalized in first-order logic. After fuzzy logic-based continuous relaxation, logical formulae are grounded onto data and neural computational graphs, hence enabling logic-induced network training. During inference, logical constraints are packaged into an iterative process and injected into the network in a form of several matrix multiplications, so as to achieve hierarchy-coherent prediction with logic reasoning. These designs together make LogicSeg a general and compact neural-logic machine that is readily integrated into existing segmentation models. Extensive experiments over four datasets with various segmentation models and backbones verify the effectiveness and generality of LogicSeg. We believe this study opens a new avenue for visual semantic parsing. Our code will be released.
TripLe: Revisiting Pretrained Model Reuse and Progressive Learning for Efficient Vision Transformer Scaling and Searching
PDF ↗One promising way to accelerate transformer training is to reuse small pretrained models to initialize the transformer, as their existing representation power facilitates faster model convergence. Previous works designed expansion operators to scale up pretrained models to the target model before training. Yet, model functionality is difficult to preserve when scaling a transformer in all dimensions at once. Moreover, maintaining the pretrained optimizer states for weights is critical for model scaling, whereas the new weights added during expansion lack these states in pretrained models. To address these issues, we propose TripLe, which partially scales a model before training, while growing the rest of the new parameters during training by copying both the warmed-up weights with the optimizer states from existing weights. As such, the new parameters introduced during training will obtain their training states. Furthermore, through serializing the scaling of model width and depth, the functionality of each expansion can be preserved. We evaluate TripLe in both single-trial model scaling and multi-trial neural architecture search (NAS). Due to the fast training convergence of TripLe, the proxy accuracy from TripLe better reveals the model quality compared to from-scratch training in multi-trial NAS. Experiments show that TripLe outperforms both from-scratch training and knowledge distillation (KD) in both training time and task performance. TripLe can also be combined with KD to achieve an even higher task accuracy. For NAS, the model obtained from TripLe outperforms DeiT-B in task accuracy with 69% reduction in parameter size and FLOPs.
Chasing Clouds: Differentiable Volumetric Rasterisation of Point Clouds as a Highly Efficient and Accurate Loss for Large-Scale Deformable 3D Registration
PDF ↗Learning-based registration for large-scale 3D point clouds has been shown to improve robustness and accuracy compared to classical methods and can be trained without supervision for locally rigid problems. However, for tasks with highly deformable structures, such as alignment of pulmonary vascular trees for medical diagnostics, previous approaches of self-supervision with regularisation and point distance losses have failed to succeed, leading to the need for complex synthetic augmentation strategies to obtain reliably strong supervision. In this work, we introduce a novel Differentiable Volumetric Rasterisation of point Clouds (DiVRoC) that overcomes those limitations and offers a highly efficient and accurate loss for large-scale deformable 3D registration. DiVRoC drastically reduces the computational complexity for measuring point cloud distances for high-resolution data with over 100k 3D points and can also be employed to extrapolate and regularise sparse motion fields, as loss in a self-training setting and as objective function in instance optimisation. DiVRoC can be successfully embedded into geometric registration networks, including PointPWC-Net and other graph CNNs. Our approach yields new state-of-the-art accuracy on the challenging PVT dataset in three different settings without training with manual ground truth: 1) unsupervised metric-based learning 2) self-supervised learning with pseudo labels generated by self-training and 3) optimisation based alignment without learning. https://github.com/mattiaspaul/ChasingClouds
Harmonizing cut-and-paste images into perceptually realistic ones is challenging, as it requires a full understanding of the discrepancies between the background of the target image and the inserted object. Existing methods mainly adjust the appearances of the inserted object via pixel-level manipulations. They are not effective in correcting color discrepancy caused by different scene illuminations and the image formation processes. We note that image colors are essentially camera ISP projection of the scene radiance. If we can trace the image colors back to the radiance field, we may be able to model the scene illumination and harmonize the discrepancy better. In this paper, we propose a novel neural approach to harmonize the image colors in a camera-independent color space, in which color values are proportional to the scene radiance. To this end, we propose a novel image unprocessing module to estimate an intermediate high dynamic range version of the object to be inserted. We then propose a novel color harmonization module that harmonizes the colors of the inserted object by querying the estimated scene radiance and re-rendering the harmonized object in the output color space. Extensive experiments demonstrate that our method outperforms the state-of-the-art approaches.