Mask DINO for MapChallenge Instance Segmentation

Installation

See installation instructions.

Pretrained Model

Download the pretrained model: maskdino_swinl_50ep_300q_hid2048_3sd1_instance_maskenhanced_mask52.3ap_box59.0ap.pth

Training

python train_net.py \
     --num-gpus 2 \
     --config-file configs/coco/instance-segmentation/swin/maskdino_R50_bs16_50ep_4s_dowsample1_2048.yaml \
     MODEL.WEIGHTS maskdino_swinl_50ep_300q_hid2048_3sd1_instance_maskenhanced_mask52.3ap_box59.0ap.pth

Evaluation

python train_net.py \
     --eval-only \
     --num-gpus 2 \
     --config-file configs/coco/instance-segmentation/swin/maskdino_R50_bs16_50ep_4s_dowsample1_2048.yaml \
     DATASETS.TEST '("satellite_test",)' \
     MODEL.WEIGHTS MODEL_CHECKPOINT_PATH

MapChallenge Instance Segmentation Results

Model	segm mAP	segm mAP@50	segm mAP@75	segm mAP_s	segm mAP_m	segm mAP_l	segm mAR	segm mAR@50	segm mAR@75
SwinS-Mask2former	0.380	0.721	0.382	0.152	0.495	0.692	0.504	0.879	0.509
SwinL-Mask2former	0.406	0.695	0.427	0.202	0.521	0.624	0.529	0.845	0.569
RTMdet-M	0.403	0.726	0.452	0.186	0.523	0.737	0.527	0.888	0.586
RTMdet-X	0.418	0.743	0.472	0.194	0.536	0.769	0.529	0.897	0.586
QueryInst-r50	0.400	0.711	0.437	0.188	0.517	0.657	0.553	0.905	0.595
QueryInst-r101	0.406	0.689	0.44	0.184	0.524	0.665	0.581	0.914	0.655
MaskDINO	0.584	0.902	0.615	0.367	0.691	0.929	0.680	0.957	0.750

MapChallenge results at >= @0.50 IOU

Model	segm mAP@50	segm mAR@50
SwinS-Mask2former	0.721	0.879
SwinL-Mask2former	0.695	0.845
RTMdet-M	0.726	0.888
RTMdet-X	0.743	0.897
QueryInst-r50	0.711	0.905
QueryInst-r101	0.689	0.914
MaskDINO	0.902	0.957

Note:

• mAP@50/mAR@50 refers to average precision/recall at 0.50 and mAP@75/mAR@75 refers to average precision/recall at 0.75 IOU
• segm mAP values are usually averaged over multiple IOUs from 0.5 to 0.95 with interval of 0.05
• In metric columns, - s = small objects - m = medium objects - l = large objects

Additional Usage

For full usage and detailed instructions, refer to Detectron2 Getting Started Guide.

Related repositories

• mmdetection training experiments