machine learning - Faster R-CNN: How to avoid multiple detections in same area? -


i use tensorflow object detection api train pascal voc dataset scratch. had on first results after 200k training steps , results okay, despite have many detections of same class in overlapping regions. example consider following detections (ignore wrong person detection in first image):

multiple detection of same motorcycle multiple detections of same aeroplane

is there general way avoid such multiple detecions of same object? guess caused overlapping region proposals detection network predicts objects fit groundtruth data above 0.7 iou threshold, maybe set threshold bit higher?

do mean literally scratch, or did use fine_tune_checkpoint?

as others have suggested, can increase iou threshold nms (decreasing .01 prevent region being detected if overlaps another).


Comments

Popular posts from this blog

python Tkinter Capturing keyboard events save as one single string -

android - InAppBilling registering BroadcastReceiver in AndroidManifest -

javascript - Z-index in d3.js -