WinARDet cuts repeated-box loops in dense vision-language grounding
Original title: WinARDet: Compressed Window Auto-Regression for Long-Horizon Vision-Language Grounding
On RSDense, authors report F1 rising from 32.16 to 39.36 and redundancy falling from 35.48% to 13.00%.
Researchers say MLLM detectors that emit coordinates autoregressively can loop on identical boxes in dense scenes because historical coordinate tokens dominate attention. WinAR compresses earlier generation windows into fixed latent tokens and keeps full token-level autoregression only in the current window. WinARDet is reported to improve RexOmni F1 on RSDense images with 200–2,000 objects and to generalize to parallel box decoding and polygon generation.
Unconfirmed: Accuracy and redundancy figures are the authors' reported results, not an independent evaluation.
- Authors trace the loop to historical-answer attention takeover under full-history autoregression.
- RSDense images contain 200 to 2,000 objects.
- Reported IoU=0.5 F1 moves from 32.16 to 39.36 versus the baseline.
- Redundancy rate falls from 35.48% to 13.00% in the authors' results.
From the original article
1 min read at huggingface.co
MLLM-based detectors formulate vision-language grounding as autoregressive coordinate generation, yet in dense scenes, they often collapse into a box repetition loop, repeatedly emitting identical boxes until the decoding budget is exhausted.
By Chenxu Wang, Hangyi Kuang, Bo-Wen Yin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.