EricChenWei/

neural-dqs-benchmark

96-variant COCO128 degradation benchmark for training and evaluating Neural Dataset Quality Score (Neural DQS) models. This dataset contains feature vectors and ground-truth mAP@0.5 values for 96 systematically degraded versions of COCO128, used to validate the hypothesis:

DQS(D) ↑ ⟹ mAP(YOLO trained on D) ↑ Result: CV Pearson r = 0.929 (n=96, p<0.001)

Each row represents one dataset variant. Features are extracted from… See the full description on the dataset page: https://huggingface.co/datasets/EricChenWei/neural-dqs-benchmark.