search

lovelyqian / CDFSOD-benchmark

A benchmark for cross-domain few-shot object detection (ECCV24 paper: Cross-Domain Few-Shot Object Detection via Enhanced Open-Set Object Detector)
Apache License 2.0
84 stars 8 forks source link

Few-shot on images with a lot of positive samples ? #29

Open tcourat opened 2 weeks ago

tcourat commented 2 weeks ago

I'd like to apply few shot object detection on images where there are plenty of expected detections (e.g a few dozens or hundreds).

The pattern is a bit repetitive, so the idea is to label only 5-10 instances, let the detector train on those instances so that it can automatically labels the remaining objects.

However it does not seem to be possible with CDFSOD if I understood well, the support images need ALL the objects to be labelled. Hence I'd have to fully annotate at least one image if I want to use it as a support image...

Correct me if I am wrong.

Thanks.

AImind commented 2 weeks ago

For a k-shot task, you only need to annotate k instances, rather than all the instances in the support images. However, since our model is pre-trained on COCO, it may not perform well in small object detection tasks.