Detectron keypoints prediction puts some annotations in wrong image.

[ArchNew]

Expected results

I expect the cyan skeletion in this image, this person.

Actual results

The Detectron's output puts the cyan skeletion in a neighboring picture...which results a very low detection rate on validation set.

Detailed steps to reproduce

I try to use detectron on posetrack's validation set [https://posetrack.net]. Unfortunately, it doesn't always put its predictions in the right picture. It sometimes puts them in a neighboring one, as I demonstrated above.

I wrote the code responsible for converting posetrack validation set's annotations into coco format. And I double checked that every image has its unique image_id. There's no two image use the same image_id. I don't think validation set's ground truth matters here, but I also double checked that the ground truth is the same before and after the conversion.

System information

• Operating system: Ubuntu 16.04
• CUDA version: 8.0
• cuDNN version: 7.2.1
• GPU models (for all devices): Titan XP
• python version: 2.7.13
• pytorch version: whatever version of caffe2 Detectron runs on.
• Anything else that seems relevant: ?

[ArchNew]

It's my bad. My conversion is wrong.