MODEL_ZOO.md

DVIS Model Zoo

Introduction

This file documents a collection of trained DVIS models. The "Config" column contains a link to the config file. Running train_net_video.py --num-gpus $num_gpus with this config file will train a model with the same setting. ResNet-50 results are trained with 8 2080Ti GPUs and Swin-L results are trained with 8 V100 GPUs.

Video Instance Segmentation

Occluded VIS

Model	Backbone	Queries	Video	AP	AP50	AP75	Config	Download
DVIS_online	R50	100	360P	30.4	54.9	29.7	yaml	baidupan|feishu
DVIS_online	R50	100	720P	30.9	55.2	30.8	yaml	baidupan|feishu
DVIS_online	SwinL(IN21k)	200	480P	46.0	70.7	48.6	yaml	baidupan|feishu
DVIS_online	SwinL(IN21k)	200	720P	47.6	72.5	50.1	yaml	baidupan|feishu
DVIS_offline	R50	100	360P	33.7	59.6	33.7	yaml	baidupan|feishu
DVIS_offline	R50	100	720P	34.6	60.1	33.3	yaml	baidupan|feishu
DVIS_offline	SwinL(IN21k)	200	480P	48.8	74.8	50.7	yaml	baidupan|feishu
DVIS_offline	SwinL(IN21k)	200	720P	50.0	75.6	52.9	yaml	baidupan|feishu

YouTubeVIS 2019

Model	Backbone	Queries	Video	AP	AP50	AP75	Config	Download
DVIS_online	R50	100	360P	51.0	72.8	56.8	yaml	baidupan|feishu
DVIS_online	SwinL(IN21k)	200	480P	63.7	86.5	70.0	yaml	baidupan|feishu
DVIS_offline	R50	100	360P	52.1	76.3	57.5	yaml	baidupan|feishu
DVIS_offline	SwinL(IN21k)	200	480P	64.8	88.2	72.6	yaml	baidupan|feishu

YouTubeVIS 2021

Model	Backbone	Queries	Video	AP	AP50	AP75	Config	Download
DVIS_online	R50	100	360P	46.3	68.2	50.6	yaml	baidupan|feishu
DVIS_online	SwinL(IN21k)	200	480P	58.6	80.8	64.6	yaml	baidupan|feishu
DVIS_offline	R50	100	360P	47.3	70.6	51.7	yaml	baidupan|feishu
DVIS_offline	SwinL(IN21k)	200	480P	60.0	82.6	68.4	yaml	baidupan|feishu

VIPSeg

Model	Backbone	Queries	Video	VPQ	VPQthing	VPQstuff	Config	Download
MinVis	R50	100	720P	-	-	-	yaml	baidupan
MinVis	SwinL(IN21k)	200	720P	-	-	-	yaml	baidupan
DVIS_online	SwinL(IN21k)	200	720P	54.7	54.8	54.6	yaml	baidupan|feishu
DVIS_offline	SwinL(IN21k)	200	720P	57.6	59.9	55.5	yaml	baidupan|feishu