A system called CapMap-MS-TTA took third place in the MUMU track of the 8th LSVOS Challenge at ECCV 2026, and its authors didn't train a single new model to get there.
The MUMU track asks one unified system to handle three jobs at once: tagging images, detecting open-vocabulary objects, and writing English captions, all under tight compute limits. CapMap-MS-TTA's answer was to chain existing tools instead of building new ones. A captioning model generates a detailed description, which then gets mapped into the challenge's official tagging vocabulary through keyword matching. Object detection runs on Florence-2, boosted with multi-scale and horizontal-flip test-time augmentation plus label-aware non-maximum suppression. That pipeline lifted the team's reproduced Florence-2 baseline score from 15.16 to 16.4815.
The result is a useful data point for anyone assuming benchmark wins require fresh training runs. Stitching together existing models with smarter inference-time tricks, no fine-tuning included, was enough to land a top-three finish against systems built for exactly this task.
The paper doesn't say how the top two teams pulled ahead, or what compute they spent doing it.