search

aim-uofa / AdelaiDet

AdelaiDet is an open source toolbox for multiple instance-level detection and recognition tasks.
https://git.io/AdelaiDet
Other
3.37k stars 646 forks source link

Getting nan for APl on both the bbox and seg results by CondInst #362

Closed zeeshanalipanhwar closed 3 years ago

zeeshanalipanhwar commented 3 years ago

Hi, I am getting results like:

[04/30 07:23:01 d2.evaluation.coco_evaluation]: Evaluation results for bbox: 
|   AP   |  AP50  |  AP75  |  APs   |  APm   |  APl  |
|:------:|:------:|:------:|:------:|:------:|:-----:|
| 17.778 | 36.530 | 15.822 | 18.069 | 12.468 |  nan  |
[04/30 07:23:01 d2.evaluation.coco_evaluation]: Some metrics cannot be computed and is shown as NaN.
.
.
.
[04/30 07:23:04 d2.evaluation.coco_evaluation]: Evaluation results for segm: 
|  AP   |  AP50  |  AP75  |  APs  |  APm  |  APl  |
|:-----:|:------:|:------:|:-----:|:-----:|:-----:|
| 0.016 | 0.126  | 0.001  | 0.016 | 0.155 |  nan  |
[04/30 07:23:04 d2.evaluation.coco_evaluation]: Some metrics cannot be computed and is shown as NaN.

How can I resolve that?

tianzhi0549 commented 3 years ago

@zeeshanalipanhwar Please check your data.

zeeshanalipanhwar commented 3 years ago

Thank you for the reply. ^^ Like what, please elaborate? The data contains instances of different sizes. I am able to visualize my annotations as well.

zeeshanalipanhwar commented 3 years ago

Here, it says, "NaN means the metric cannot be computed (e.g. due to no predictions made).". What I have done is use a pretrained set of weights and train it for 3k iterations, so maybe that's why the large objects are not learned well.

zeeshanalipanhwar commented 3 years ago

I just trained it for 90k iterations, still got nan for APL. How are the sizes for s, m, and l instances decided?