search

amirbar / DETReg

Official implementation of the CVPR 2022 paper "DETReg: Unsupervised Pretraining with Region Priors for Object Detection".
https://amirbar.net/detreg
Apache License 2.0
336 stars 46 forks source link

Reproducing the Results of Table 3 & Table 4 #18

Closed mmaaz60 closed 2 years ago

mmaaz60 commented 3 years ago

Hi,

How can I reproduce the results of Table 3 which is about Few-shot detection performance for the 20 novel categories on the COCO dataset.

Also, in Table 4 (Comparison to semi-supervised detection methods), it is mentioned in the paper that you pretrained the network on the entire coco train2017 unlabeled images and then fine-tuned on X% of data. But the instructions in README or the corresponding config files load the ImageNet100 pretrained weights for fine-tuning on COCO. Kindly guide me, what process may I follow to reproduce the results reported in Table 4.

Thanks

amirbar commented 2 years ago

Hi,

  1. Table 3 - I currently don't plan to release the Few-Shot code base (dataloaders, classes, datasplit, etc.) which is based on this repo.
  2. To pretrain on COCO you should set the --dataset param to coco_pretrain ( see https://github.com/amirbar/DETReg/blob/main/main.py#L366). Also, here is the COCO pretraining model checkpoint if this helps. Then, to finetune you need to set the relevant hyperparam described in the paper (epochs, lr, lr drop).
usama-axcelerate commented 1 year ago
  1. Table 3 - I currently don't plan to release the Few-Shot code base (dataloaders, classes, datasplit, etc.) which is based on this repo.

Have you open sourced the few shot code yet?