Low Confidence Scores for High IoU Segments in THUMOS14 Predictions

[Damon-Salvatore-liu]

Hello,

Thank you very much for your outstanding work！

Now I encountered an issue while working with your codebase on the THUMOS14 dataset. For example, in video_test_0000004, I noticed that the predicted segment overlaps well with the ground truth segment, achieving a high IoU (e.g., 0.616060). However, the confidence score for this prediction is unexpectedly low (e.g., 0.06995).

This issue is not limited to this video; I observed similar behavior in other videos as well.

Could you please provide insights into why this might happen? Have you analyzed this phenomenon before? I’d also like to invite the community to discuss potential reasons and solutions for this.

Thank you for your time and support!

[tzzcl]

a video will have multiple groundtruth actions, what you see in the table is that the predictions may have low IoU with specific actions [0.2, 1.1], but it may have high overlap with other groundtruth actions.

[Damon-Salvatore-liu]

Thank you very much for your response.
However, you might have misunderstood my point. What I meant is that the predicted segments have relatively high IoU values with the ground-truth labels, but their classification scores are relatively low (as shown in the red box). I would like to ask if this is normal. Theoretically, if the IoU score is high, the classification score should also be relatively high.