A standard hardware bottleneck when training deep neural networks is GPU memory. The bulk of memory is occupied by caching intermediate tensors for gradient computation in the backward pass. We propose a novel method to reduce this footprint - Dropping Intermediate Tensors (DropIT). DropIT drops min-k elements of the intermediate tensors and approximates gradients from the sparsified tensors in the backward pass. Theoretically, DropIT reduces noise on estimated gradients and therefore has a higher rate of convergence than vanilla-SGD. Experiments show that we can drop up to 90\% of the intermediate tensor elements in fully-connected and convolutional layers while achieving higher testing accuracy for Visual Transformers and Convolutional Neural Networks on various tasks (e.g., classification, object detection, instance segmentation). Our code and models are available at https://github.com/chenjoya/dropit.
Scene graph generation (SGG) in satellite imagery (SAI) benefits promoting understanding of geospatial scenarios from perception to cognition. In SAI, objects exhibit great variations in scales and aspect ratios, and there exist rich relationships between objects (even between spatially disjoint objects), which makes it attractive to holistically conduct SGG in large-size very-high-resolution (VHR) SAI. However, there lack such SGG datasets. Due to the complexity of large-size SAI, mining triplets <subject, relationship, object> heavily relies on long-range contextual reasoning. Consequently, SGG models designed for small-size natural imagery are not directly applicable to large-size SAI. This paper constructs a large-scale dataset for SGG in large-size VHR SAI with image sizes ranging from 512 x 768 to 27,860 x 31,096 pixels, named STAR (Scene graph generaTion in lArge-size satellite imageRy), encompassing over 210K objects and over 400K triplets. To realize SGG in large-size SAI, we propose a context-aware cascade cognition (CAC) framework to understand SAI regarding object detection (OBD), pair pruning and relationship prediction for SGG. We also release a SAI-oriented SGG toolkit with about 30 OBD and 10 SGG methods which need further adaptation by our devised modules on our challenging STAR dataset. The dataset and toolkit are available at: https://linlin-dev.github.io/project/STAR.
How to effectively fuse cross-modal information is the key problem for RGB-D salient object detection. Early fusion and the result fusion schemes fuse RGB and depth information at the input and output stages, respectively, hence incur the problem of distribution gap or information loss. Many models use the feature fusion strategy but are limited by the low-order point-to-point fusion methods. In this paper, we propose a novel mutual attention model by fusing attention and contexts from different modalities. We use the non-local attention of one modality to propagate long-range contextual dependencies for the other modality, thus leveraging complementary attention cues to perform high-order and trilinear cross-modal interaction. We also propose to induce contrast inference from the mutual attention and obtain a unified model. Considering low-quality depth data may detriment the model performance, we further propose selective attention to reweight the added depth cues. We embed the proposed modules in a two-stream CNN for RGB-D SOD. Experimental results have demonstrated the effectiveness of our proposed model. Moreover, we also construct a new challenging large-scale RGB-D SOD dataset with high-quality, thus can both promote the training and evaluation of deep models.
Perception systems in modern autonomous driving vehicles typically take inputs from complementary multi-modal sensors, e.g., LiDAR and cameras. However, in real-world applications, sensor corruptions and failures lead to inferior performances, thus compromising autonomous safety. In this paper, we propose a robust framework, called MetaBEV, to address extreme real-world environments involving overall six sensor corruptions and two extreme sensor-missing situations. In MetaBEV, signals from multiple sensors are first processed by modal-specific encoders. Subsequently, a set of dense BEV queries are initialized, termed meta-BEV. These queries are then processed iteratively by a BEV-Evolving decoder, which selectively aggregates deep features from either LiDAR, cameras, or both modalities. The updated BEV representations are further leveraged for multiple 3D prediction tasks. Additionally, we introduce a new M2oE structure to alleviate the performance drop on distinct tasks in multi-task joint learning. Finally, MetaBEV is evaluated on the nuScenes dataset with 3D object detection and BEV map segmentation tasks. Experiments show MetaBEV outperforms prior arts by a large margin on both full and corrupted modalities. For instance, when the LiDAR signal is missing, MetaBEV improves 35.5% detection NDS and 17.7% segmentation mIoU upon the vanilla BEVFusion model; and when the camera signal is absent, MetaBEV still achieves 69.2% NDS and 53.7% mIoU, which is even higher than previous works that perform on full-modalities. Moreover, MetaBEV performs fairly against previous methods in both canonical perception and multi-task learning settings, refreshing state-of-the-art nuScenes BEV map segmentation with 70.4% mIoU.
Locating 3D objects from a single RGB image via Perspective-n-Point (PnP) is a long-standing problem in computer vision. Driven by end-to-end deep learning, recent studies suggest interpreting PnP as a differentiable layer, allowing for partial learning of 2D-3D point correspondences by backpropagating the gradients of pose loss. Yet, learning the entire correspondences from scratch is highly challenging, particularly for ambiguous pose solutions, where the globally optimal pose is theoretically non-differentiable w.r.t. the points. In this paper, we propose the EPro-PnP, a probabilistic PnP layer for general end-to-end pose estimation, which outputs a distribution of pose with differentiable probability density on the SE(3) manifold. The 2D-3D coordinates and corresponding weights are treated as intermediate variables learned by minimizing the KL divergence between the predicted and target pose distribution. The underlying principle generalizes previous approaches, and resembles the attention mechanism. EPro-PnP can enhance existing correspondence networks, closing the gap between PnP-based method and the task-specific leaders on the LineMOD 6DoF pose estimation benchmark. Furthermore, EPro-PnP helps to explore new possibilities of network design, as we demonstrate a novel deformable correspondence network with the state-of-the-art pose accuracy on the nuScenes 3D object detection benchmark. Our code is available at https://github.com/tjiiv-cprg/EPro-PnP-v2.
Tracking and following objects of interest is critical to several robotics use cases, ranging from industrial automation to logistics and warehousing, to healthcare and security. In this paper, we present a robotic system to detect, track, and follow any object in real-time. Our approach, dubbed ``follow anything'' (FAn), is an open-vocabulary and multimodal model -- it is not restricted to concepts seen at training time and can be applied to novel classes at inference time using text, images, or click queries. Leveraging rich visual descriptors from large-scale pre-trained models (foundation models), FAn can detect and segment objects by matching multimodal queries (text, images, clicks) against an input image sequence. These detected and segmented objects are tracked across image frames, all while accounting for occlusion and object re-emergence. We demonstrate FAn on a real-world robotic system (a micro aerial vehicle) and report its ability to seamlessly follow the objects of interest in a real-time control loop. FAn can be deployed on a laptop with a lightweight (6-8 GB) graphics card, achieving a throughput of 6-20 frames per second. To enable rapid adoption, deployment, and extensibility, we open-source all our code on our project webpage at https://github.com/alaamaalouf/FollowAnything . We also encourage the reader to watch our 5-minutes explainer video in this https://www.youtube.com/watch?v=6Mgt3EPytrw .
Annotation errors are a challenge not only during training of machine learning models, but also during their evaluation. Label variations and inaccuracies in datasets often manifest as contradictory examples that deviate from established labeling conventions. Such inconsistencies, when significant, prevent models from achieving optimal performance on metrics such as mean Average Precision (mAP). We introduce the notion of "label convergence" to describe the highest achievable performance under the constraint of contradictory test annotations, essentially defining an upper bound on model accuracy. Recognizing that noise is an inherent characteristic of all data, our study analyzes five real-world datasets, including the LVIS dataset, to investigate the phenomenon of label convergence. We approximate that label convergence is between 62.63-67.52 mAP@[0.5:0.95:0.05] for LVIS with 95% confidence, attributing these bounds to the presence of real annotation errors. With current state-of-the-art (SOTA) models at the upper end of the label convergence interval for the well-studied LVIS dataset, we conclude that model capacity is sufficient to solve current object detection problems. Therefore, future efforts should focus on three key aspects: (1) updating the problem specification and adjusting evaluation practices to account for unavoidable label noise, (2) creating cleaner data, especially test data, and (3) including multi-annotated data to investigate annotation variation and make these issues visible from the outset.
In autonomous systems, sensor calibration is essential for a safe and efficient navigation in dynamic environments. Accurate calibration is a prerequisite for reliable perception and planning tasks such as object detection and obstacle avoidance. Many existing LiDAR calibration methods require overlapping fields of view, while others use external sensing devices or postulate a feature-rich environment. In addition, Sensor-to-Vehicle calibration is not supported by the vast majority of calibration algorithms. In this work, we propose a novel target-based technique for extrinsic Sensor-to-Sensor and Sensor-to-Vehicle calibration of multi-LiDAR systems called CaLiV. This algorithm works for non-overlapping FoVs, as well as arbitrary calibration targets, and does not require any external sensing devices. First, we apply motion to produce FoV overlaps and utilize a simple unscented Kalman filter to obtain vehicle poses. Then, we use the Gaussian mixture model-based registration framework GMMCalib to align the point clouds in a common calibration frame. Finally, we reduce the task of recovering the sensor extrinsics to a minimization problem. We show that both translational and rotational Sensor-to-Sensor errors can be solved accurately by our method. In addition, all Sensor-to-Vehicle rotation angles can also be calibrated with high accuracy. We validate the simulation results in real-world experiments. The code is open source and available on https://github.com/TUMFTM/CaLiV.
Vision Transformers have witnessed prevailing success in a series of vision tasks. However, these Transformers often rely on extensive computational costs to achieve high performance, which is burdensome to deploy on resource-constrained devices. To alleviate this issue, we draw lessons from depthwise separable convolution and imitate its ideology to design an efficient Transformer backbone, i.e., Separable Vision Transformer, abbreviated as SepViT. SepViT helps to carry out the local-global information interaction within and among the windows in sequential order via a depthwise separable self-attention. The novel window token embedding and grouped self-attention are employed to compute the attention relationship among windows with negligible cost and establish long-range visual interactions across multiple windows, respectively. Extensive experiments on general-purpose vision benchmarks demonstrate that SepViT can achieve a state-of-the-art trade-off between performance and latency. Among them, SepViT achieves 84.2% top-1 accuracy on ImageNet-1K classification while decreasing the latency by 40%, compared to the ones with similar accuracy (e.g., CSWin). Furthermore, SepViT achieves 51.0% mIoU on ADE20K semantic segmentation task, 47.9 AP on the RetinaNet-based COCO detection task, 49.4 box AP and 44.6 mask AP on Mask R-CNN-based COCO object detection and instance segmentation tasks.
Fast and robust object grasping in clutter is a crucial component of robotics. Most current works resort to the whole observed point cloud for 6-Dof grasp generation, ignoring the guidance information excavated from global semantics, thus limiting high-quality grasp generation and real-time performance. In this work, we show that the widely used heatmaps are underestimated in the efficiency of 6-Dof grasp generation. Therefore, we propose an effective local grasp generator combined with grasp heatmaps as guidance, which infers in a global-to-local semantic-to-point way. Specifically, Gaussian encoding and the grid-based strategy are applied to predict grasp heatmaps as guidance to aggregate local points into graspable regions and provide global semantic information. Further, a novel non-uniform anchor sampling mechanism is designed to improve grasp accuracy and diversity. Benefiting from the high-efficiency encoding in the image space and focusing on points in local graspable regions, our framework can perform high-quality grasp detection in real-time and achieve state-of-the-art results. In addition, real robot experiments demonstrate the effectiveness of our method with a success rate of 94% and a clutter completion rate of 100%. Our code is available at https://github.com/THU-VCLab/HGGD.
Segmentation evaluation metrics traditionally rely on binary decision logic: predictions are either correct or incorrect, based on rigid IoU thresholds. Detection--based metrics such as F1 and mAP determine correctness at the object level using fixed overlap cutoffs, while overlap--based metrics like Intersection over Union (IoU) and Dice operate at the pixel level, often overlooking instance--level structure. Panoptic Quality (PQ) attempts to unify detection and segmentation assessment, but it remains dependent on hard-threshold matching--treating predictions below the threshold as entirely incorrect. This binary framing obscures important distinctions between qualitatively different errors and fails to reward gradual model improvements. We propose SoftPQ, a flexible and interpretable instance segmentation metric that redefines evaluation as a graded continuum rather than a binary classification. SoftPQ introduces tunable upper and lower IoU thresholds to define a partial matching region and applies a sublinear penalty function to ambiguous or fragmented predictions. These extensions allow SoftPQ to exhibit smoother score behavior, greater robustness to structural segmentation errors, and more informative feedback for model development and evaluation. Through controlled perturbation experiments, we show that SoftPQ captures meaningful differences in segmentation quality that existing metrics overlook, making it a practical and principled alternative for both benchmarking and iterative model refinement.
We propose a simple yet effective proposal-based object detector, aiming at detecting highly-overlapped instances in crowded scenes. The key of our approach is to let each proposal predict a set of correlated instances rather than a single one in previous proposal-based frameworks. Equipped with new techniques such as EMD Loss and Set NMS, our detector can effectively handle the difficulty of detecting highly overlapped objects. On a FPN-Res50 baseline, our detector can obtain 4.9\% AP gains on challenging CrowdHuman dataset and 1.0\% $\text{MR}^{-2}$ improvements on CityPersons dataset, without bells and whistles. Moreover, on less crowed datasets like COCO, our approach can still achieve moderate improvement, suggesting the proposed method is robust to crowdedness. Code and pre-trained models will be released at https://github.com/megvii-model/CrowdDetection.
Knowledge distillation transfers knowledge from the teacher network to the student one, with the goal of greatly improving the performance of the student network. Previous methods mostly focus on proposing feature transformation and loss functions between the same level's features to improve the effectiveness. We differently study the factor of connection path cross levels between teacher and student networks, and reveal its great importance. For the first time in knowledge distillation, cross-stage connection paths are proposed. Our new review mechanism is effective and structurally simple. Our finally designed nested and compact framework requires negligible computation overhead, and outperforms other methods on a variety of tasks. We apply our method to classification, object detection, and instance segmentation tasks. All of them witness significant student network performance improvement. Code is available at https://github.com/Jia-Research-Lab/ReviewKD
We introduce R-Sparse R-CNN, a novel pipeline for oriented ship detection in Synthetic Aperture Radar (SAR) images that leverages sparse learnable proposals enriched with background contextual information, termed background-aware proposals (BAPs). The adoption of sparse proposals streamlines the pipeline by eliminating the need for proposal generators and post-processing for overlapping predictions. The proposed BAPs enrich object representation by integrating ship and background features, allowing the model to learn their contextual relationships for more accurate distinction of ships in complex environments. To complement BAPs, we propose Dual-Context Pooling (DCP), a novel strategy that jointly extracts ship and background features in a single unified operation. This unified design improves efficiency by eliminating redundant computation inherent in separate pooling. Moreover, by ensuring that ship and background features are pooled from the same feature map level, DCP provides aligned features that improve contextual relationship learning. Finally, as a core component of contextual relationship learning in R-Sparse R-CNN, we design a dedicated transformer-based Interaction Module. This module interacts pooled ship and background features with corresponding proposal features and models their relationships. Experimental results show that R-Sparse R-CNN delivers outstanding accuracy, surpassing state-of-the-art models by margins of up to 12.8% and 11.9% on SSDD and RSDD-SAR inshore datasets, respectively. These results demonstrate the effectiveness and competitiveness of R-Sparse R-CNN as a robust framework for oriented ship detection in SAR imagery. The code is available at: www.github.com/ka-mirul/R-Sparse-R-CNN.
The advancement of computer vision and machine learning has made datasets a crucial element for further research and applications. However, the creation and development of robots with advanced recognition capabilities are hindered by the lack of appropriate datasets. Existing image or video processing datasets are unable to accurately depict observations from a moving robot, and they do not contain the kinematics information necessary for robotic tasks. Synthetic data, on the other hand, are cost-effective to create and offer greater flexibility for adapting to various applications. Hence, they are widely utilized in both research and industry. In this paper, we propose the dataset HabitatDyn, which contains both synthetic RGB videos, semantic labels, and depth information, as well as kinetics information. HabitatDyn was created from the perspective of a mobile robot with a moving camera, and contains 30 scenes featuring six different types of moving objects with varying velocities. To demonstrate the usability of our dataset, two existing algorithms are used for evaluation and an approach to estimate the distance between the object and camera is implemented based on these segmentation methods and evaluated through the dataset. With the availability of this dataset, we aspire to foster further advancements in the field of mobile robotics, leading to more capable and intelligent robots that can navigate and interact with their environments more effectively. The code is publicly available at https://github.com/ignc-research/HabitatDyn.
Different from Object Detection, Visual Grounding deals with detecting a bounding box for each text-image pair. This one box for each text-image data provides sparse supervision signals. Although previous works achieve impressive results, their passive utilization of annotation, i.e. the sole use of the box annotation as regression ground truth, results in a suboptimal performance. In this paper, we present SegVG, a novel method transfers the box-level annotation as Segmentation signals to provide an additional pixel-level supervision for Visual Grounding. Specifically, we propose the Multi-layer Multi-task Encoder-Decoder as the target grounding stage, where we learn a regression query and multiple segmentation queries to ground the target by regression and segmentation of the box in each decoding layer, respectively. This approach allows us to iteratively exploit the annotation as signals for both box-level regression and pixel-level segmentation. Moreover, as the backbones are typically initialized by pretrained parameters learned from unimodal tasks and the queries for both regression and segmentation are static learnable embeddings, a domain discrepancy remains among these three types of features, which impairs subsequent target grounding. To mitigate this discrepancy, we introduce the Triple Alignment module, where the query, text, and vision tokens are triangularly updated to share the same space by triple attention mechanism. Extensive experiments on five widely used datasets validate our state-of-the-art (SOTA) performance.
As moving objects always draw more attention of human eyes, the temporal motive information is always exploited complementarily with spatial information to detect salient objects in videos. Although efficient tools such as optical flow have been proposed to extract temporal motive information, it often encounters difficulties when used for saliency detection due to the movement of camera or the partial movement of salient objects. In this paper, we investigate the complimentary roles of spatial and temporal information and propose a novel dynamic spatiotemporal network (DS-Net) for more effective fusion of spatiotemporal information. We construct a symmetric two-bypass network to explicitly extract spatial and temporal features. A dynamic weight generator (DWG) is designed to automatically learn the reliability of corresponding saliency branch. And a top-down cross attentive aggregation (CAA) procedure is designed so as to facilitate dynamic complementary aggregation of spatiotemporal features. Finally, the features are modified by spatial attention with the guidance of coarse saliency map and then go through decoder part for final saliency map. Experimental results on five benchmarks VOS, DAVIS, FBMS, SegTrack-v2, and ViSal demonstrate that the proposed method achieves superior performance than state-of-the-art algorithms. The source code is available at https://github.com/TJUMMG/DS-Net.
The "Roaring 20s" of visual recognition began with the introduction of Vision Transformers (ViTs), which quickly superseded ConvNets as the state-of-the-art image classification model. A vanilla ViT, on the other hand, faces difficulties when applied to general computer vision tasks such as object detection and semantic segmentation. It is the hierarchical Transformers (e.g., Swin Transformers) that reintroduced several ConvNet priors, making Transformers practically viable as a generic vision backbone and demonstrating remarkable performance on a wide variety of vision tasks. However, the effectiveness of such hybrid approaches is still largely credited to the intrinsic superiority of Transformers, rather than the inherent inductive biases of convolutions. In this work, we reexamine the design spaces and test the limits of what a pure ConvNet can achieve. We gradually "modernize" a standard ResNet toward the design of a vision Transformer, and discover several key components that contribute to the performance difference along the way. The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt. Constructed entirely from standard ConvNet modules, ConvNeXts compete favorably with Transformers in terms of accuracy and scalability, achieving 87.8% ImageNet top-1 accuracy and outperforming Swin Transformers on COCO detection and ADE20K segmentation, while maintaining the simplicity and efficiency of standard ConvNets.
Accurate segmentation of organelle instances, e.g., mitochondria, is essential for electron microscopy analysis. Despite the outstanding performance of fully supervised methods, they highly rely on sufficient per-pixel annotated data and are sensitive to domain shift. Aiming to develop a highly annotation-efficient approach with competitive performance, we focus on weakly-supervised domain adaptation (WDA) with a type of extremely sparse and weak annotation demanding minimal annotation efforts, i.e., sparse point annotations on only a small subset of object instances. To reduce performance degradation arising from domain shift, we explore multi-level transferable knowledge through conducting three complementary tasks, i.e., counting, detection, and segmentation, constituting a task pyramid with different levels of domain invariance. The intuition behind this is that after investigating a related source domain, it is much easier to spot similar objects in the target domain than to delineate their fine boundaries. Specifically, we enforce counting estimation as a global constraint to the detection with sparse supervision, which further guides the segmentation. A cross-position cut-and-paste augmentation is introduced to further compensate for the annotation sparsity. Extensive validations show that our model with only 15% point annotations can achieve comparable performance as supervised models and shows robustness to annotation selection.
We propose a method to efficiently equip the Segment Anything Model (SAM) with the ability to generate regional captions. SAM presents strong generalizability to segment anything while is short for semantic understanding. By introducing a lightweight query-based feature mixer, we align the region-specific features with the embedding space of language models for later caption generation. As the number of trainable parameters is small (typically in the order of tens of millions), it costs less computation, less memory usage, and less communication bandwidth, resulting in both fast and scalable training. To address the scarcity problem of regional caption data, we propose to first pre-train our model on objection detection and segmentation tasks. We call this step weak supervision pretraining since the pre-training data only contains category names instead of full-sentence descriptions. The weak supervision pretraining allows us to leverage many publicly available object detection and segmentation datasets. We conduct extensive experiments to demonstrate the superiority of our method and validate each design choice. This work serves as a stepping stone towards scaling up regional captioning data and sheds light on exploring efficient ways to augment SAM with regional semantics. The project page, along with the associated code, can be accessed via https://xk-huang.github.io/segment-caption-anything/.