Search

Date Range
Advanced
Adjust the alpha value (0 to 1)
Adjust the relevance threshold (0 to 1)
arxiv.org Tutorial

[2610.25090] Think Visually Across Space and Time: Grounded Reasoning for Video Segmentation

A grounding-based approach augments video reasoning with temporally anchored primitives to bridge language and visual evidence, enabling accurate video segmentation through supervised and reinforcement learning, with public release plans.

1 min read Listed 28 days ago in Hugging Face Daily Papers
github.com Documentation

GitHub - AndreiiArhire/LightMIS: LightMIS: Ultra-Lightweight Medical Image Segmentation Without a Stage-Wise Decoder · GitHub

LightMIS is a compact, multi-scale medical image segmentation model that replaces traditional stage-wise decoders with a Scale-Aligned Projection head and an Adaptive Fusion Cascade, yielding competitive accuracy with far fewer parameters and lower compute. It is evaluated on multiple datasets under nnU-Net protocol and demonstrates strong efficiency on mobile GPUs.

2 mins read Listed 12 days ago in reddit/MachineLearning/LightMIS: Ultra-lightweight medical image segmentation for edge devices [R]
annals-csis.org Tutorial

Annals of Computer Science and Information Systems, Volume 43

A multi-stage U-Net based method improves retinal vessel segmentation across varying image resolutions, using coarse-to-fine processing with multiple scales, validated on four public datasets, and released as open-source.

1 min read Listed 25 days ago in Bluesky
github.com Tools and repositories

GitHub - AlbertRG99/VehiclePartsSegmentatorFTModels: Car-part segmentation on VehicleSeg10K: YOLO26x-seg, RF-DETR Seg 2XL, DINOv3 + decoder and SAM 3 zero-shot, trained and scored with one shared evaluation (mIoU, mAP, part F1, left/right confusion, speed, VRAM). Best: DINOv3 at 91.5 mIoU. · GitHub

The repository compares YOLO26x-seg, RF-DETR Seg 2XL, DINOv3, and zero-shot SAM 3 for segmenting 13 car parts on VehicleSeg10K. DINOv3 achieves the best mean IoU at 91.5%, while SAM 3 has difficulty distinguishing parts on the car’s left and right sides.

8 mins read Listed 2 days ago in reddit/computervision/Vehicle Part Segmentation | Training and Model Comparison
www.r-bloggers.com Tutorial

Semantic segmentation in rplatypus and pyplatypus: what is in the box | R-bloggers

The article explains how to configure semantic segmentation training in rplatypus and pyplatypus using YAML or constructor arguments. It covers data formats, model settings, training options, and validation errors, with examples in R and Python.

13 mins read Listed 3 days ago in RSS: R-bloggers
link.springer.com Tutorial

A systematic survey of open-ended visual segmentation: taxonomies, challenges and future directions | Artificial Intelligence Review | Springer Nature Link

The document surveys open-ended visual segmentation, outlining OSS, OVS, and OWS as key directions, presenting a unified framework, datasets, metrics, and method taxonomies, and discussing challenges and future research avenues. It also provides performance comparisons and notes public availability of resources.

2 mins read Listed 20 days ago in RSS: Latest Results
huggingface.co Tools and repositories

MrBr1ghts1d3/VehiclePartsSegmentatorFTModels · Hugging Face

This repository contains three models fine-tuned to segment 13 car parts, with evaluation results, training details, and usage examples. DINOv3 scores highest on mIoU, while YOLO26x-seg has the lowest reported inference time.

4 mins read Listed 2 days ago in reddit/computervision/Vehicle Part Segmentation | Training and Model Comparison
arxiv.org Tutorial

[2610.24452] HumanSkill-4D: Training-Free Language Querying of Animatable Gaussian Avatars in Space and Time

The paper presents HumanSkill-4D, a training-free framework for language-driven avatar segmentation that uses a shared multimodal model and geometric operators to select Gaussian regions corresponding to body parts and garments based on language, pose, or motion. It introduces a semantic memory per avatar, programs combining operators to verify query conditions on motion data, and benchmarks showing improved part, 3D, and 4D referring segmentation with efficient setup.

1 min read Listed 29 days ago in Hugging Face Daily Papers
huggingface.co Tools and repositories

thedeoxen/MOSS-Image-segmentation-controlnet · Hugging Face

This repository provides a segmentation Control LoRA for MOSS-Image-SFT that uses flat-color region maps to preserve image structure while prompts set the subjects and visual details. It includes ComfyUI setup instructions, recommended sampling settings, and tips for preparing control maps.

3 mins read Listed 4 days ago in reddit/StableDiffusion/MOSS-Image-SFT control LoRAs — pose, depth, canny and the rest
arxiv.org Tutorial

[2610.26872] Commanding at the Speed of Sound: Audio-Centric Referring Expression Segmentation for Aerial Scenes

The paper shifts UAV referring expression segmentation from text-centric to audio-centric by using raw audio to guide pixel-level localization, introducing the Aerial-AudioRES benchmark and the ARES-Net model, which aligns acoustic signals with visual features and achieves strong accuracy with reduced latency.

2 mins read Listed 24 days ago in Bluesky