START.md

Getting started with IAI

Model zoo

YTVIS2019

Name	Backbone	Aug	Pretrain	AP	AP50	AP75	AR1	AR10	Model
IAI+CondInst	R50		COCO	39.2	61.3	43.5	39.2	46.5	gdrive
IAI+CondInst	R50	ms	COCO	39.9	62.3	43.9	40.1	46.8	gdrive
IAI+CondInst	R101	ms	COCO	43.7	67.2	48.4	41.7	50.0	gdrive

YTVIS2021

Name	Backbone	Pretrain	AP	AP50	AP75	AR1	AR10	Model
IAI+CondInst	R50	COCO	38.0	59.1	43.0	34.8	44.5	gdrive

OVIS

Name	Backbone	Pretrain	AP	AP50	AP75	AR1	AR10	Model
IAI+CondInst	R50	COCO+YTVIS2019	20.6	38.9	20.3	11.9	25.8	gdrive

Training

Before training, please put the COCO pretrained model under the model folder.

To train model with multple GPUs, run:

bash tools/dist_train.sh $CONFIG_PATH 2

For example, to train IAI+CondInst r50 on YouTube-VIS 2019 with 2 GPUs, run:

bash tools/dist_train.sh configs/iai/ytvis2019_iai_condinst_r50.py 2

After training, the model will be saved in the output folder.

Inference & Evaluation

Evaluating on YouTube-VIS 2019 r50 with single GPU, run:

python tools/test.py $CONFIG_PATH $MODEL_PATH --eval segm

Evaluating on YouTube-VIS 2019 r50 with multiple GPUs (e.g. 2 GPUs), run:

bash tools/dist_test.sh $CONFIG_PATH $MODEL_PATH 2 

After evaluating, a json file with the predicted result will be generated as output/results.json. To get validataion results, please zip the json file and upload it to the codalab server for YouTube-VIS 2019, YouTube-VIS 2021 and OVIS.