Extremely low P value while training on BDD100K

[naruarjun]

Hello,
Thank you for the great work on this project. While training on BDD100K, the Precision value for object detection I get is extremely low.

Driving area Segment: Acc(0.967)    IOU (0.832)    mIOU(0.897)
Lane line Segment: Acc(0.626)    IOU (0.250)  mIOU(0.617)
Detect: P(0.049)  R(0.885)  [email protected](0.728)  [email protected]:0.95(0.391)

Am I doing something wrong? The other values match the values given in the readme, but Precision is not given so I don't have a reference.

[naruarjun]

I have found the reason, it is because the code reports P and R for the highest recall IOU. I believe it would be more beneficial to report the highest F1 IOU.

[Riser6]

I have found the reason, it is because the code reports P and R for the highest recall IOU. I believe it would be more beneficial to report the highest F1 IOU.
Yeah, you are right! We shall consider your suggestion! Thanks for your attention to our project!

[RahulRewale]

Hi @naruarjun @Riser6

Where is this logic?

I am getting below results after training for 200 epochs on BDD100k dataset (13 classes):

Driving area Segment: Acc(0.968)    IOU (0.830)    mIOU(0.896)
Lane line Segment: Acc(0.585)    IOU (0.462)  mIOU(0.724)
Detect: P(0.769)  R(0.193)  [email protected](0.177)  [email protected]:0.95(0.094)

As you can see, I am getting very low recall and mAP values. Do you know what could be wrong?

[s0966066980]

Hi @naruarjun @Riser6

Where is this logic?

I am getting below results after training for 200 epochs on BDD100k dataset (13 classes):

Driving area Segment: Acc(0.968)    IOU (0.830)    mIOU(0.896)
Lane line Segment: Acc(0.585)    IOU (0.462)  mIOU(0.724)
Detect: P(0.769)  R(0.193)  [email protected](0.177)  [email protected]:0.95(0.094)

As you can see, I am getting very low recall and mAP values. Do you know what could be wrong?

Did you solve it? I am using your 13-class YOLOP.

[RahulRewale]

@s0966066980
As per my experiments, not using the focal loss (setting _C.LOSS.FL_GAMMA to 0 in default.py) improves recall value and decreases precision.

When I set it to 1.5 or 2.0, recall was < 0.2, and when I set it to 0.0, recall was ~ 0.32. This is still not good, but since the authors haven't shared the results, I don't know what should be the target value.