[Request] Non Maximum Suppression

[BonJovi1]

Hi AntonMu,

System information

• What is the top-level directory of the model you are using: TrainYourOwnYolo/3_Inference
• Have I written custom code (as opposed to using a stock example script provided in the repo): No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MacOS High Sierra
• TensorFlow version (use command below): 1.15.2
• CUDA/cuDNN version: None
• GPU model and memory: None, training on my CPU
• Exact command to reproduce: python Detector.py

Describing the problem

Thanks a lot for the amazing code and documentation, I was able to successfully train the YOLO model on a custom dataset. However while testing/inference, my model detects multiple bounding boxes around the same object. So it says there are two faces even though there is only one. I looked up online and learnt that this problem can be solved using Non Maximum Suppression. I tried implementing it, with the idea that we could only print the bbox with maximum confidence score, but I was unable to do so (I'm not exactly understanding how to use the box_list_list variable). Could you kindly let me know how exactly could we modify your code to achieve this?

Thanks a lot,
Abhinav

[AntonMu]

Hi @BonJovi1

You are welcome. There is a related issue here: #124 <- NMS is already part of the code (it is a core component of YOLO).

Do your boxes have similar confidence levels? I recommend trying different confidence thresholds with e.g.

python Detector.py --confidence 0.5

and other thresholds. Generally, the higher the threshold the fewer boxes you will get.

[BonJovi1]

Hi @AntonMu
Thanks a bunch! I'm sorry, I should've seen the confidence flag feature, somehow missed that. It still wasn't good enough for me though, because I was still getting multiple detections with confidence scores around 0.6.
So I just manually iterated over out_boxes and only picked up those whose out_score was the highest, and that did the job for me! I guess I should've tried it first properly before raising the issue!

Thanks a lot,
Abhinav