Abstract Video Object Segmentation (VOS) is a key component in computer vision applications, including surveillance, autonomous driving, and robotics. However, existing VOS models often struggle with generalization to new videos with complex, topologically transforming deformable objects (eg. cooking, assembling, state change), degraded environments and long video sequences, resulting in tracking drift, low recall and memory saturation. We developed Mu ltiple object VOS and tracking S mart Mem ory architecture (MuSMem), a generalizable approach that incorporates three key innovations: (i) fusing SAM with High-Quality masks alongside appearance-based candidate-selection to refine coarse segmentation masks, resulting in improved object boundaries; (ii) dynamic smart memory that manages a history of key frames based on a novel information preserving gain , combined with relevance and freshness spatio-temporal criteria; and (iii) explores the use of monocular depth maps for occlusion robustness. MuSMem significantly reduces memory usage, reduces drift, tracks complex object topological changes and improves long-term prediction performance. MuSMem can be integrated with Vision-Language Models (VLMs) for zero-shot generalization to unseen visual domains. Experiments using VOS benchmark datasets show that MuSMem ranks first on VOTSt-2024, Long Video Dataset and LVOS, and second on VOTS-2024, demonstrating the best generalizability and state-of-the-art performance across single-, multi-, and complex VOS tasks.

Hypoxic Ischemic Encephalopathy (HIE) represents a brain dysfunction, affecting approximately 1 to 5 per 1000 full-term neonates. The precise delineation and segmentation of HIE-related lesions in neonatal brain Magnetic Resonance Images (MRI) are pivotal in advancing outcome predictions, identifying patients at high risk, elucidating neurological manifestations, and assessing treatment efficacies. Despite its importance, the development of algorithms for segmenting HIE lesions from MRI volumes has been impeded by data scarcity. Addressing this critical gap, we organized the first BONBID-HIE challenge with diffusion MRI data (Apparent Diffusion Coefficient (ADC) maps) for HIE lesion segmentation, in conjunction with the MICCAI 2023. Totally 14 algorithms were submitted, employing a gamut of cutting-edge automatic machine-learning-based segmentation algorithms. Our comprehensive analysis of HIE lesion segmentation and submitted algorithms facilitates an in-depth evaluation of the current technological zenith, outlines directions for future advancements, and highlights persistent hurdles. To foster ongoing research and benchmarking, the annotated HIE dataset, developed algorithm dockers, and unified evaluation codes are accessible through a dedicated online platform (https://bonbid-hie2023.grand-challenge.org).

From content moderation to wildlife conservation, the number of applications that require models to recognize nuanced or subjective visual concepts is growing. Traditionally, developing classifiers for such concepts requires substantial manual effort measured in hours, days, or even months to identify and annotate data needed for training. Even with recently proposed Agile Modeling techniques, which enable rapid bootstrapping of image classifiers, users are still required to spend 30 minutes or more of monotonous, repetitive data labeling just to train a single classifier. Drawing on Fiske's Cognitive Miser theory, we propose a new framework that alleviates manual effort by replacing human labeling with natural language interactions, reducing the total effort required to define a concept by an order of magnitude: from labeling 2,000 images to only 100 plus some natural language interactions. Our framework leverages recent advances in foundation models, both large language models and vision-language models, to carve out the concept space through conversation and by automatically labeling training data points. Most importantly, our framework eliminates the need for crowd-sourced annotations. Moreover, our framework ultimately produces lightweight classification models that are deploy-able in cost-sensitive scenarios. Across 15 subjective concepts and across 2 public image classification datasets, our trained models outperform traditional Agile Modeling as well as state-of-the-art zero-shot classification models like ALIGN, CLIP, CuPL, and large visual question answering models like PaLI-X.

Abstract Discriminating salient moving objects against complex, cluttered backgrounds, with occlusions and challenging environmental conditions like weather and illumination, is essential for stateful scene perception in autonomous systems. We propose a novel deep architecture, named DeepFTSG, for robust moving object detection that incorporates single and multi-stream multi-channel USE-Net trellis asymmetric encoders extending U-Net with squeeze and excitation (SE) blocks and a single shared decoder network for fusing multiple motion and appearance cues. DeepFTSG is a deep learning based approach that builds upon our previous hand-engineered flux tensor split Gaussian (FTSG) change detection video analysis algorithm which won the CDNet CVPR Change Detection Workshop challenge competition. DeepFTSG generalizes much better than top-performing motion detection deep networks, such as the scene-dependent ensemble-based FgSegNet_v2, while using an order of magnitude fewer weights. Short-term motion and longer-term change cues are estimated using general-purpose unsupervised methods—flux tensor and multi-modal background subtraction, respectively. DeepFTSG was evaluated using the CDnet-2014 change detection challenge dataset, the largest change detection video sequence benchmark with 12.3 billion labeled pixels, and had an overall F-measure of 97%. We also evaluated the cross-dataset generalization capability of DeepFTSG trained solely on CDnet-2014 short video segments and then evaluated on unseen SBI-2015, LASIESTA and LaSOT benchmark videos. On the unseen SBI-2015 dataset, DeepFTSG had an F-measure accuracy of 87%, more than 30% higher compared to the top-performing deep network FgSegNet_v2 and outperforms the recently proposed KimHa method by 17%. On the unseen LASIESTA, DeepFTSG had an F-measure of 88% and outperformed the best recent deep learning method BSUV-Net2.0 by 3%. On the unseen LaSOT with axis-aligned bounding box ground-truth, network segmentation masks were converted to bounding boxes for evaluation, DeepFTSG had an F-Measure of 55%, outperforming KimHa method by 14% and FgSegNet_v2 by almost 1.5%. When a customized single DeepFTSG model is trained in a scene-dependent manner for comparison with state-of-the-art approaches, then DeepFTSG performs significantly better, reaching an F-Measure of 97% on SBI-2015 (+ 10%) and 99% on LASIESTA (+ 11%). The source code, pre-trained weights, and video demo for DeepFTSG are available at https://github.com/CIVA-Lab/DeepFTSG .