search

tensorflow / models

Models and examples built with TensorFlow
Other
76.94k stars 45.79k forks source link

Custom definitions for small, medium, large boxes in object detection #8375

Closed tispratik closed 4 years ago

tispratik commented 4 years ago

We're training our object detection based model on 1280x720 images which are being resized to 300x300. Most of our bounding boxes have an area greater than 96^96 which is the maximum size for medium boxes. So almost all of our boxes lie in the large bucket.

Is there a way I can set custom sizes for these definitions so that we can have more useful mAP, precision, recall values on tensorboard?

Currently, we get no values for small and medium sized boxes.

Screen Shot 2020-04-08 at 6 24 38 PM
sainisanjay commented 4 years ago

Hi @tispratik are you able to solve your issue? as i am facing same issue.

tispratik commented 4 years ago

The framework uses pycocotools. For now I have the pycocotools code and changed the values to what I need. The values are based on the original image height and width. Search for class Params in cocoeval.py.

sainisanjay commented 4 years ago

Thanks @tispratik since my image size is 1820x940 and most of annotated object fall under large [Car, Bus, Van, Truck, Pedestrian]so i can simply removed small and medium from areaRng and check only mAP for large. Is my thought is right?

tispratik commented 4 years ago

The thought is right, but if some sizes of objects are not training well you wouldn't know if you don't have metrics breakdown by the different sizes.

pkulzc commented 4 years ago

That's right, this requires modifying pycocotools. So you can modify your own pycocotools repo to fit your use cases.

IamExperimenting commented 2 years ago

@tispratik, thanks for sharing this. Currently, I'm training an object detection model, I'm just wondering how to visualise the mAP value during training and evaluation. Now, I see only the losses for training in the tensorboard but not the mAP value. During evaluation I can see a single pointfor mAP all, small, medium, large. how to visualise the mAP value similar to your image which you have posted above. Could you please provide any article to replicate the same? It would be really helpful.

tispratik commented 2 years ago

Hi @IamExperimenting, it could mean that none of your bboxes meet the size criteria for mAP small/medium/large... check your bbox sizes.