Seeking Advice: Improving Detection When Animal Is Absent in Most Frames

[skyshi0]

Hi SLEAP team and community,

I’m seeking advice on improving my model’s ability to avoid false positives when my animal is not present in most frames.

Setup:

• Video resolution: 1080×1080 px, ~1px per mm
• Animal size:18-25 px length
• Animal presence: ~15% of total frames

The bottom-up model performs okay when the target is present and visible. However, since I want to use the detection timestamps for downstream behavioral analysis, I need the model to also correctly identify when the animal is absent (i.e., predict no instances).

Problem

Initially, I trained the model based on ~1200 labeled frames only when the target is present. The bottom-up model seems to have fewer false positive frames compared to top-down and single models when I ran for an entire video(visually inspected, I haven’t systematically quantified the false positive frames). So I was thinking whether I could improve its performance by adding more ‘invisible frames’(meaning add those frames as labeled frames but set all the nodes to invisible), but after I added 133 ‘invisible frames’, the bottom-up model seems to not improve and still identified random locations as the target.

Question

Do you have any advice on how to improve model training when:(1)The target is very small (~19 px),(2)The target only appears in a small subset of frames (~15%), (3)And I want to minimize false positives when the target is absent?

I’m wondering:

1. Should I increase or decrease the ratio of “invisible frames” in the training dataset or does it matter?
2. . Are there recommended data augmentation or preprocessing steps for such small targets?
3. Is the resolution too low to identify the target? I know it doesn't meet the criteria for mouse ~2mm per pixel, but I wonder if it can still be improved since I don't need the fine features of the target.

Any suggestions or examples from similar projects would be greatly appreciated!
Thanks in advance for your time and insights!