detection

DenseCL: Transferring to Detection

The train_net.py script reproduces the object detection experiments on Pascal VOC and COCO.

Instruction

1. Install detectron2.

   $ git clone https://github.com/facebookresearch/detectron2.git
   $ cd detectron2
   $ git checkout 3e71a2711bec
   $ python -m pip install -e .
   

   This requires cuda10.2 to work.

2. Convert a pre-trained model to detectron2's format:

   python3 convert-pretrain-to-detectron2.py input.pth.tar output.pkl
   

3. Put dataset under "./datasets" directory, following the directory structure requried by detectron2.

      $ mkdir -p datasets && cd datasets
      $ ln -s VOC2007 .
      $ ln -s VOC2012 .
   

4. Run training:

   # r50 
   python train_net.py --config-file configs/pascal_voc_R_50_C4_24k_moco.yaml \
    --num-gpus 8 MODEL.WEIGHTS ./output.pkl
   # r101 
   python train_net.py --config-file configs/pascal_voc_R_101_C4_24k_moco.yaml \
    --num-gpus 8 MODEL.WEIGHTS ./output.pkl
   

   Or you can see dist_train.sh for the training scripts.

Results

Below are the results on Pascal VOC 2007 test, fine-tuned on 2007+2012 trainval for 24k iterations using Faster R-CNN with a R50/R101-C4 backbone:

pretrain	AP50	AP	AP75
ImageNet-1M, R50, supervised	81.3	53.5	58.8
ImageNet-1M, R50, MoCo v1, 200ep	81.5	55.9	62.6
ImageNet-1M, R50, MoCo v2, 200ep	82.4	57.0	63.6
ImageNet-1M, R50, MoCo v2, 800ep	82.5	57.4	64.0
ImageNet-1M, R50, DenseCL, 200ep	82.7	58.5	65.6
ImageNet-1M, R101, DenseCL, 200ep	83.57	61.02	68.20
ImageNet-1M, R50, RegionCL-D, 200ep	83.32	58.72	65.57
ImageNet-1M, R101, RegionCL-D, 200ep	84.30	61.59	68.17

Note: These results are means of 5 trials. Variation on Pascal VOC is large: the std of AP50, AP, AP75 is expected to be 0.2, 0.2, 0.4 in most cases. We recommend to run 5 trials and compute means.

denseCL, r50:
82.64/58.32/64.60
82.64/58.41/64.89

denseCL, r101:
83.57/61.02/68.20
83.52/60.89/67.32

regionCL-D, r50:
83.24/58.84/65.98
83.40/58.60/65.16

regionCL-D, r101:
84.22/61.48/68.14
84.39/61.70/68.21