Object Detection

We use Mask-RCNN to evaluate the down-stream performance. Please refer to the paper for detailed settings.

Requirements

First follow the installation guidance in the main readme file to prepare the environment. For object detection, additional packages shall be installed:

• mmcv (1.x.0) installation guidance
• mmdet (1.x.0) installation guidance Other version may also work.

Results

Scale	Model	Box AP	Mask AP	checkpoint
Micro	U-HaloNet	40.3	37.3	Download
	U-PVT	35.9	34.2	Download
	U-Swin Transformer	36.6	34.6	Download
	U-ConvNeXt	39.2	36.4	Download
	U-InternImage	39.5	36.6	Download
Tiny	U-HaloNet	46.9	42.4	Download
	U-PVT	44.2	40.6	Download
	U-Swin Transformer	44.3	40.5	Download
	U-ConvNeXt	44.3	40.5	Download
	U-InternImage	47.2	42.5	Download
Small	U-HaloNet	48.2	43.3	Download
	U-PVT	46.1	41.9	Download
	U-Swin Transformer	46.4	42.1	Download
	U-ConvNeXt	45.6	41.2	Download
	U-InternImage	47.8	43.0	Download
Base	U-HaloNet	49.0	43.8	Download
	U-PVT	46.4	42.3	Download
	U-Swin Transformer	47.0	42.2	Download
	U-ConvNeXt	46.7	42.2	Download
	U-InternImage	48.7	43.8	Download

Usage

Training

• To train a mask rcnn based on unified-swin with slurm
  • remember to modify the path to the pertained ckpt in ./configs/unified_models/

# MODEL_TYPE: halonet, pvt, swin, convnext
# SCALE: micro, tiny, small, base
bash shells/train.py [MODEL_TYPE] [SCALE]

• For training on a single machine, run the following command:

bash shells/dist_train.py [CONFIG] [NUM_GPUS] --auto-scale-lr

Evaluation

• To test the trained model with slurm

bash shells/test.py [MODEL_TYPE] [SCALE]

• For single machine evaluation, run the following command:

bash shells/dist_test.py [CONFIG] [CKPT_PATH] [NUM_GPUS]