Instructions on training about who_where

[fwtan]

I will try to provide a training script (hopefully soon...)
Basically, you may need to provide the inpainted image, instead of the original color image, as the input for training. We include an Adobe script to automatically in-paint the COCO images via Photoshop.
The inpainted images should be placed in the COCO data directory, as indicated in the inpainted_path_from_indices function in coco.py.
The cfg.TOP_K is a parameter controlling the number of locations and shapes of the bounding boxes sampled from the model. Like in the paper, we evaluated our top-8 results, which are produced from the combinations of the top 2 location predictions, top 2 size
predictions and top 2 retrieved segments.

[zhengpingzhou]

Thanks for your response! I finally successfully ran training by ignoring that variable and removing the np.repeat... statements in generator.

Besides, as mentioned in your answer, inpainting is done by Adobe Photoshop. Is there any reason for not using open source implementations such as context encoder or generative inpainting?

[fwtan]

We choose to use Photoshop as it can handle masks with arbitrary shapes and achieve reasonable visual quality.
But feel free to try other methods such as Globally and Locally Consistent Image Completion and generative inpainting.
There is of course NO reason for not using these open source methods if they can help achieve better results!