Event
MV2GF Multi-View Pedestrian Detection Method Released as Open Source
1 reports · 1 sources
The ECCV'26 paper proposes and open-sources the multi-view pedestrian detection method MV2GF, which replaces perspective projection with multi-view features and 3D point maps from the pretrained visual geometry foundation model DA3 to aggregate BEV features; the author team states its purpose is to improve cross-view generalization. Under the setting where training and test camera configurations differ, MV2GF's MODA on GMVD-D, MVPerception, and Wildtrack is 4.6, 4.7, and 2.2 points higher than the previous best, respectively; however, when facing older methods trained on the same camera configuration as the test configuration, its detection performance is still slightly inferior. The cost is speed: the author team froze DA3 and trained only the remaining small modules, measuring end-to-end inference at 3.3 FPS on 6 views with an A100 80GB, lower than the 4.5–5.2 FPS of previous methods.
Report timeline
- ECCV'26 open-sources MV2GF: using a visual geometry foundation model to improve cross-view pedestrian detection generalization, with MODA up by as much as 4.7 points
An ECCV'26 paper proposes MV2GF, a multi-view pedestrian detection method that uses the multi-view features and 3D point maps of the pretrained visual geometry foundation model DA3 to aggregate BEV features instead of perspective projection. Under the setting where training and test camera configurations differ, MV2GF achieves MODA on GMVD-D, MVPerception, and Wildtrack that is 4.6, 4.7, and 2.2 points higher than the previous best, respectively.
3D视觉工坊 · 微信公众号 ·