detection

Transfer Learning: Object Detection

Running

1. Install Detectron2.

2. Convert pre-trained models to Detectron2 models:

python convert_pretrained.py model.pth det_model.pkl

3. Set up data folders following Detectron2's datasets instruction.

4. Go to Detectron2's folder, and run:

python tools/train_net.py \
  --num-gpus 8 \
  --config-file /path/to/config/config.yaml \
  MODEL.WEIGHTS /path/to/model/det_model.pkl

where config.yaml is the config file listed under the configs folder.

Results

(1) Results with ResNet-50 (200 epochs pre-training)

Mask-RCNN, FPN:

Pretrain	Arch	Detector	lr sched	box AP	mask AP	download
No (rand init.)	R-50	Mask-RCNN, FPN	1x	32.8	29.9	config  |  model  |  log
Supervised	R-50	Mask-RCNN, FPN	1x	39.7	35.9	config  |  model  |  log
InstDis	R-50	Mask-RCNN, FPN	1x	38.8	35.2	config  |  model  |  log
PIRL	R-50	Mask-RCNN, FPN	1x	38.6	35.1	config  |  model  |  log
MoCo v1	R-50	Mask-RCNN, FPN	1x	39.4	35.6	config  |  model  |  log
InfoMin Aug.	R-50	Mask-RCNN, FPN	1x	40.6	36.7	config  |  model  |  log
No (rand init.)	R-50	Mask-RCNN, FPN	2x	38.4	34.7	config  |  model  |  log
Supervised	R-50	Mask-RCNN, FPN	2x	41.6	37.6	config  |  model  |  log
InstDis	R-50	Mask-RCNN, FPN	2x	41.3	37.3	config  |  model  |  log
PIRL	R-50	Mask-RCNN, FPN	2x	41.2	37.4	config  |  model  |  log
MoCo v1	R-50	Mask-RCNN, FPN	2x	41.7	37.5	config  |  model  |  log
MoCo v2	R-50	Mask-RCNN, FPN	2x	41.7	37.6	config  |  model  |  log
InfoMin Aug.	R-50	Mask-RCNN, FPN	2x	42.5	38.4	config  |  model  |  log
No (rand init.)	R-50	Mask-RCNN, FPN	6x	42.7	38.6	config  |  model  |  log
Supervised	R-50	Mask-RCNN, FPN	6x	42.6	38.5	config  |  model  |  log
InfoMin Aug.	R-50	Mask-RCNN, FPN	6x	43.6	39.2	config  |  model  |  log

Mask-RCNN, C4:

Pretrain	Arch	Detector	lr sched	box AP	mask AP	download
Supervised	R-50	Mask-RCNN, C4	1x	38.2	33.3	config
MoCo	R-50	Mask-RCNN, C4	1x	38.5	33.6	config
InfoMin Aug.	R-50	Mask-RCNN, C4	1x	39.0	34.1	config
Supervised	R-50	Mask-RCNN, C4	2x	40.0	34.7	config
MoCo	R-50	Mask-RCNN, C4	2x	40.7	35.6	config
InfoMin Aug.	R-50	Mask-RCNN, C4	2x	41.3	36.0	config

(2) Results with other architecture

See paper.

TODO: add logs and pre-trained detector models