liming-ai/AlignDet

Hi,how can i get proposals with cluster info in detr

Minorant opened this issue · 1 comments

Hi, thanks for your impressive work!
I'm trying to run DETR on the VOC dataset, but I notice that only coco's post-clustering proposal is provided(train2017_ratio3size0008@0.5_cluster-id-as-class.json). So how can I generate my own post clustering proposal? Is there any code available or more detailed steps for this please, thanks!

Hi, thanks for your impressive work! I'm trying to run DETR on the VOC dataset, but I notice that only coco's post-clustering proposal is provided(train2017_ratio3size0008@0.5_cluster-id-as-class.json). So how can I generate my own post clustering proposal? Is there any code available or more detailed steps for this please, thanks!

Hi, thanks for your interest in AlignDet~

Since AlignDet is pre-trained on COCO, so we only provide the clustered proposals on COCO. When fine-tuning the pre-trained model on VOC, we have the ground truth annotations.

Unfortunately, we cannot release the code for the clustering proposal, but this procedure should be simple:

  1. Running the selective search algorithm (Please refer to SoCo) to obtain unsupervised proposals.
  2. Extracting the features of these proposals with RoI Align operation (Can be done with supervised/unsupervised backbone)
  3. Performing PCA algorithm to reduce the dimension of features (We use the dimension of 256, but others should also be fine)
  4. Using the K-means algorithm to cluster these proposals.

Thanks again for your question, I hope it can help you~