Be careful when reading the output mask from Annotation tool

[sandonair007]

Motivation

Train a segmentation model on Platform, when dataset annotation is from ABEJA Annotation Tool.

Problem

Generally we use combined mask (multiple colors exist in one png image) as the target. Though we have pre-defined label-color map dict for the annotation job, mixed colors often appear along the margin of different color blocks, which results in unknown label ids when converting png to label array using PIL.Image.Palette. Finally the training ends up with errors like:

2019-10-09T19:56:12+09:00: /pytorch/aten/src/THCUNN/SpatialClassNLLCriterion.cu:103: void cunn_SpatialClassNLLCriterion_updateOutput_kernel(T *, T *, T *, long *, T *, int, int, int, int, int, long) [with T = float, AccumT = float]: block: [1,0,0], thread: [194,0,0] Assertion `t >= 0 && t < n_classes` failed.

Example

The above image shows the original mask.

The white dots denote the pixels where mix color exists (new label id will be created.)

Solutions

• Treat the edges as background. This approach will not affect the result too much as the amount of such pixels is small compared to color blocks.
• Use nearest-neighbor to map the edge color to one of the pre-defined color.

[keigohtr]

I have confirmed the label image Annotation-tool outputs includes the mixed color because some label's area are over wrapped.

This situation is unnatural for a general segmentation dataset. I think they, Annotation-tool, should fix this.

[toshilow]

I researched about combined logic in annotation tool. it looks like the layers are correctly overcoated.

In Annotation Screen shot

Convined image data

Zoom up image by drawing tool (Pixelmator Pro)

[sandonair007]

Thank you for the update. My example is from https://abeja-datalake-production.s3.amazonaws.com/3fa7-1888808746696/20191007/012627-0cebbeb0-d4d9-48ab-9518-293f14ff25a8?AWSAccessKeyId=ASIASVXDTXMFUXFWVE7T&Signature=XzQ0PiWeaUOxMe4GAuzJs2axo6o%3D&x-amz-security-token=AgoJb3JpZ2luX2VjEMX%2F%2F%2F%2F%2F%2F%2F%2F%2F%2FwEaDmFwLW5vcnRoZWFzdC0xIkcwRQIhAI%2B34S5fSN7u2hium9Jkul%2FnTdk3IVQX49vjyQ17wAORAiAddrf0qjES%2FsHvaVMwBw1xfKhvsmRK9oSttUF6c4jt9yqvAgjO%2F%2F%2F%2F%2F%2F%2F%2F%2F%2F8BEAEaDDE4NDA4NzE5ODQ3NSIMbr%2B%2Bib2sQbBqUzDbKoMCt0wfWgvueSwZyEanElGDwPVS1xEiOR1nXzkqN9BjeSFtMDpRJ189RcJZ7KAEXWJZd2wor6QmSDIt6j7SZpc8gCUBXXRniZp8TkaR8%2F655n9EpAGszAtDjV%2BYZJh3UmPiI9dy86JkawhT8RPnl%2FtfGzHBPaUsPkvYTiXBZ0rB7XpiBFpAAZLWoN9%2BKmaQ30rKIB8Dsrn4Fr1lqh4VNSMjO12MydQPhEaX%2BujGBk2J8auAAKz7P9OEx66P%2BTBiU56jTFT8cNNxomlc0lcC2ngo3JRClQK4zXhFjKTNFTibKnKDNRixGf%2F0RnUjiqrAKP6s7L8LW6DkYAYHECrrT%2B8DZsWumjD40OntBTq0AcDu7jiBvsjL5u7LckIW%2FmJcBSZkM0oSLv9sOqhdzJD5D2VBzBWNZ%2Bt40ooBHpuCNDsrxjWJG5BzLAGnf8Z06Z%2BhzVY8odqcQDxYkIQTiGIDajn8nKZ7mhKMq9TL3hn%2Bd09orS7klK%2BQl%2FMuZiX9LT8tJVw5XFfKUrPOihXA0nStk9%2BdH3wNAowgU9TcwzjoE1ZY2z88kjK%2Fg1nyO6QDg6AMbaDV5JVMtqPx9mzIep8EToBk9Q%3D%3D&Expires=1572510555

| organization_id | 1815063019817 | | datalake_channel_id | 1888808746696 | | file_id | 20191007T012627-0cebbeb0-d4d9-48ab-9518-293f14ff25a8 |