search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.77k stars 7.96k forks source link

half object and right labelling #6525

Open haviduck opened 4 years ago

haviduck commented 4 years ago

hey, im labelling alooot of fish under water. now im struggling with the correct way of labelling a bunch of fish behind eachother. the closest ones i can see the entire fish and thats easy, but others i only get the tail part.

so my current approach was labbeling whole fish and complete fish as the same object. the result being that 1 fish can get detected as 2 since i labelled the tail as the same as a fish.

other times the fish swim by so i thlight, hey i gotta tag that aswell. we did the same with pytorch and getting some bad results there aswell with instance segmentation etc.

its alot of photos to tag so before i start over now i wanted to hear what you guys recommend? ive seen some projects only labelling the head, but thats not accurate enough for my purpose.

so: TLDR: Should i only label whole objects and darknet Will understand that a part of the object is infact the correct object?

stephanecharette commented 4 years ago

This page is a bit simplistic (I meant it more as a beginner's guide) but here it is anyway: https://www.ccoderun.ca/darkmark/ImageMarkup.html#PartialObjectMarkup

If you can be certain that a tail or a head is a certain type of fish, then I'd label it that way. Don't create a new class for partial fish, or fish tail, etc..., unless you need to identify that and you mean to mark every tail, etc.

The problem you may run into is if you have a cluster of different fish together, all semi-obscured or too small to label each one individually. What I did for a car project a while ago was to create a separate class I called "many vehicles". If the picture had lots of cars parked along the side of the road, or on a highway, then the ones in the distance that couldn't be individually marked up would be grouped together as a mass of "many vehicles"...or "many fish". That way you still get recognition of objects, though it wont tell you which kind.

image

haviduck commented 4 years ago

good read, thank you! reason im struggling with partials is cause i count objects using deep sort, and when 1 fish is 3 it lowers my accuracy. but alot of good advice there! gonna tune my labels and retrain, perhaps i can escape some labelhours. thanks a bunch for replying, helps alot