Hi everyone,
I’m an independent researcher (undergraduate, Intelligent Robotics program)
looking for someone eligible to endorse my first submission to arXiv’s
cs.CV category.
Paper: “Auto Dataset Builder: An Automated Annotation Framework with
Vision-Language Verification and Neural Dataset Quality Scoring”
Abstract (short version): We present ADB, a framework that converts
unlabeled images into YOLO-compatible annotated datasets via a three-stage
pipeline (YOLOv11 proposals → SAM2 mask refinement → CLIP/Vision-LLM
zero-shot verification), requiring zero manual annotation. We also
introduce Neural DQS, a learned dataset-quality score that predicts
post-training mAP@0.5 from 6 dataset-level features without training a
model (CV Pearson r=0.929 on COCO128 degradation variants). On a COCO2017
motorcycle subset, ADB’s fully automatic annotations reach 95.6% of
manually-annotated mAP@0.5, and one active-learning round closes the gap.
Preprint (Zenodo DOI, includes link to code): Auto Dataset Builder: An Automated Annotation Framework with Vision-Language Verification and Neural Dataset Quality Scoring
If you’ve submitted 3+ papers to cs.CV (or related CS categories) and are
willing to help, here’s my endorsement link:
Thanks so much for considering — happy to answer any questions about the
work!
arxiv、computer-vision、endorsement