video_id:{
objects:{
1: {category: ..., frames: ...}
2: {category: ..., frames: ...}
}
}
Every 'objects' term is a dict for a video, in which the keys of object categories always start from '1' for different videos and there is no object id term at all. Besides, I do not follow the idea that the object ids are equal to the pixel values in the annotation images, since there is no object id and the colors in different images representing different object categories are same (like they are all red in the annotation images).
Posted by: zombder @ July 24, 2018, 8:23 a.m.Hi! The 'object id' here does not mean the unique id for object over the whole data set. Instead, the value of 'object id' is the same as the pixel values in the segmentation PNG files. For example, if 'object id' is 1, all the pixels equals 1 in the corresponding PNG file are the segmentation of the object.
Posted by: fyc0624 @ July 24, 2018, 6:56 p.m.Got it! I misunderstood it. Sorry.
Posted by: zombder @ July 25, 2018, 2:22 a.m.