Performance difference in tensorflow

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )

http://pjreddie.com/darknet/

Other

21.8k stars 7.97k forks source link

Performance difference in tensorflow #4890

Open aditya-droid opened 4 years ago

aditya-droid commented 4 years ago

Hi AlexeyAB,

While testing out various models such as yolov3 and yolov3-spp as well as yolov2 I found that there is a significant performance difference between inference accuracies of darknet and TensorFlow.

Could you guide me towards a method by which I can improve the accuracy in TensorFlow as well? Is it related to image preprocessing or other steps that are executed before inferencing with the model in darknet?

AlexeyAB commented 4 years ago

What is the difference? Show FPS / mAP for both cases.

pragyakale commented 4 years ago

@aditya-droid what function are you using for NMS? I'm also building yolov4 in tensorflow and it seems that tf.image.combined_non_max_suppression works on CPU which increases the inference time.