0

I'm using nvidia Transfer Learning Toolkit to detect cars in some video frames.

I found some dataset (for example https://www.jpjodoin.com/urbantracker/dataset.html and https://www.kaggle.com/aalborguniversity/aau-rainsnow) and I noticed that usually parked cars are not labeled, and covered under a mask.

Why shouldn't I add also their labels? It would be easy to label them because they are static objects and I could copy-paste in all labels set. So why in video dataset they are not labelled?

Andre Goulart
  • 854
  • 2
  • 25
Francesco Pagani
  • 11
  • 1

1 Answers1

1
  1. Model architecture:

In machine learning, static image detectors can be is very different from video detectors, as movement plays a big role on the task. So, even when comparing frames the objects are similar, when digesting the video, a model can learn very different things. Maybe adding parked cars to the database increased false positives, mistakenly labeling other static noises as cars.

  1. Business goal:

Why are you labeling cars from security camera, in the first place? What is your goal here? Maybe if you want to know car density in a parking lot, than labeling parked cars is very useful. But if you just need to know traffic flow, than parked cars will be just a distraction, noise from the data. So maybe whoever built the dataset had a different goal in mind.

Andre Goulart
  • 854
  • 2
  • 25
  • On further reflection, I noticed that there is a black mask aviable covering area of "no interest", including parked cars. I'm going to use something similar: the user select the "important" area and the model will detect all cars inside. I just need some YOLO behavior, so my business goal is to find all cars. However, including parked ones may over-represent one red Ferrari that will be labeled in all frames. – Francesco Pagani Aug 18 '21 at 10:44