Training on grayscale has poor performance on color

[daskandalis]

I have very unequal performance on the same video when it's in grayscale or in color (trained in grayscale, worse in color). I have two copies of the same video, one loaded in grayscale and the other in color. The model training_config.json suggests grayscale will be enforced:

"preprocessing": {
            "ensure_rgb": false,
            "ensure_grayscale": true,
            ...
        }

Is this expected? Is there an option to force the conversion to grayscale during inference?

[daskandalis]

Toggling grayscale in the sleap-label interface is sufficient. When the videos are on grayscale mode, they're tracked well, when they're in color mode, the tracking is poor.

[Marielamend]

Hello @daskandalis,

Yes, selecting the toggle feature under the videos tab converts a video/all videos in a project into grayscale if not already.

There is also a way of converting it under Input Data options in the training pipeline, but this seems to not convert all videos in a project into grayscale in some cases. I can put this under an issue, and our team can further look into it. Thank you for your inquiry!

Best,
Mariela