search

IDEA-Research / T-Rex

[ECCV2024] API code for T-Rex2: Towards Generic Object Detection via Text-Visual Prompt Synergy
https://deepdataspace.com/home
Other
1.98k stars 120 forks source link

Could you show me two test samples in FSC147? #21

Closed Hzzone closed 3 months ago

Hzzone commented 5 months ago
image image

1121.jpg and 7621.jpg are dominant in the final MAE of FSC-147 dataset. For example, there is a gap of MAE=3000/1190 without '1123.jpg'. I cannot achieve a good result using your demo.

Mountchicken commented 5 months ago

Hi @Hzzone Sry for the late reply. I tried these two images and T-Rex indeed performs bad. Here is the reason: both two images have objects that are far more than the maximun number(900) that T-Rex can detect. Below, I set the score threshold to 0. and T-Rex can detect 900 boxes, which is still far not enough. image (5)