ProCal is an inference-time proposal calibration method for open-vocabulary object detection. The method combines a localization-aware foreground score with a background-aware suppression score to form a proposal prior, improving the localization quality of classification scores from a frozen vision-language model. The authors report that, when applied to CLIPSelf ViT-L/14, ProCal improves APr by 2.5 on OV-LVIS. Their analysis attributes the gain to proposal-level reranking that suppresses false novel activations on background regions and ranks true novel proposals above background and partial-object proposals.
No heat snapshots are available in the last 24 hours.