Video Instance Segmentation

Overview

VIDEO INSTANCE SEGMENTATION (SkyVIS)

The SkyVIS challenge aims at simultaneous segmentation and tracking of the instances of objects of interest in videos. Given an input video, the output of this task consists of the segmentation masks of all instances of a predefined list of categories and the instance identities across frames (time) to be associated.

Metrics

We use similar metrics as used in YouTubeVIS challenge. The main difference is that, we compute the AR values not only for small, medium and large objects but also tiny and micro sized objects. For the definition of those object sizes, please check the dataset page.

FILE FORMAT FOR SUBMISSION

 

 [{           
    'video_id'       : int, 
    'category_id'    : int, 
    'segmentations'  : [RLE or None],
    'score'          : float,
 }]