search

xiaofeng94 / VL-PLM

Exploiting unlabeled data with vision and language models for object detection, ECCV 2022
86 stars 7 forks source link

The code for the semi-supervised experiment #11

Open fcjian opened 1 year ago

fcjian commented 1 year ago

Where are the code and model for the semi-supervised experiment?

fcjian commented 1 year ago

@xiaofeng94 How did you set the score thresh to filter the bbox from CLIP and the nms iou thresh to merge the bboxes from CLIP and STAC for the semi-supervised experiment?

xiaofeng94 commented 1 year ago

Hey, sorry for the late reply.

For semi-supervised experiment, we use STAC code with minor modifications to adopt our PLs. For thresholds, set them to keep number of boxes in a reasonable range. In our paper, we keep around 4 pseudo labels per image.

fcjian commented 1 year ago

The author set the score threshold to 0.9 in STAC paper while you set the score threshold to 0.8 in OVD experiment. I want to know what threshold you set in your semi-supervised experiment?

xiaofeng94 commented 1 year ago

Hey, if I remembered correctly, we used a threshold of 0.8 for semi-supervised experiments the same as OVD.