Human Segmentation Tool

Upload an image or video, choose a model and task.

Running on CPU. Image inference: 1–15 s depending on model. Video and live modes use GASNet only (fast enough for per-frame inference).

Select Test Type
Select Model

GASNet v4 — latest checkpoint. MobileNetV3-Small backbone with ARFB multi-scale geometry, GARB_Geo geometry injection, and LKA attention. Single forward pass, no detection stage. Best for real-time use.

Select Task

Original image with detected person highlighted in green.

Example Images

Training data: LIP · COCO-person · MADS · Penn-Fudan
GitHub: Human Segmentation Evaluation