search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.7k stars 7.96k forks source link

Model false detection on total different scenario #7530

Open www7890 opened 3 years ago

www7890 commented 3 years ago

Hi all, thanks to this wonderful github, I have trained a model based on DVR driving scenario and it perform pretty well. While I detected my office around my seat with the model, I found out that lots of object, such as chairs, PCs ,telephones, are recognized as car or bus or some other class.

My question is:

  1. Does this mean my model is terrible? seriously overfitting?
  2. Should the false detection on some other scenario be solved or it is not the big deal?

Thanks!

shubham-shahh commented 3 years ago

What are the classes of your network? what was the average aspect ratio of the images used for training and what is your network's resolution?

www7890 commented 3 years ago

What are the classes of your network? what was the average aspect ratio of the images used for training and what is your network's resolution?

Classes are cars, person, truck, moto-bike and bus Img-size would be 512*512, 720p

shubham-shahh commented 3 years ago

What are the classes of your network? what was the average aspect ratio of the images used for training and what is your network's resolution?

Classes are cars, person, truck, moto-bike and bus Img-size would be 512*512, 720p

and model resolution?

shubham-shahh commented 3 years ago

Hi all, thanks to this wonderful github, I have trained a model based on DVR driving scenario and it perform pretty well. While I detected my office around my seat with the model, I found out that lots of object, such as chairs, PCs ,telephones, are recognized as car or bus or some other class.

My question is:

  1. Does this mean my model is terrible? seriously overfitting?
  2. Should the false detection on some other scenario be solved or it is not the big deal?

Thanks!

This problem often occurs when the model is trained only with the images containing the object occupying the entire frame, for example, if you train a model for license plate detection, and your images contain magnified license plates and not a real scenario where it is attached to a car.

www7890 commented 3 years ago

magnified

Oh, I see. So it is all about the size images are trained in model. My source images size are 1080p, and model resolution is 512*512, with multi-scale skill 1/1.5 to 1.5 I have checked my data after feeding to model, there is sometimes huge object because they are very near to the camera.

I am curious how to solve this problem, if I don't use the multi-scale skill, the map of model will drop about 4 %, it seems that my model 'can't see' the object smaller than 4_4 pixels with 512_512 size (about 20*20 pixels for source images 1080p)

Hi all, thanks to this wonderful github, I have trained a model based on DVR driving scenario and it perform pretty well. While I detected my office around my seat with the model, I found out that lots of object, such as chairs, PCs ,telephones, are recognized as car or bus or some other class. My question is:

  1. Does this mean my model is terrible? seriously overfitting?
  2. Should the false detection on some other scenario be solved or it is not the big deal?

Thanks!

This problem often occurs when the model is trained only with the images containing the object occupying the entire frame, for example, if you train a model for license plate detection, and your images contain magnified license plates and not a real scenario where it is attached to a car.

Oh, I see. So it is all about the size images are trained in model. My source images size are 1080p, and model resolution is 512*512, with multi-scale skill 1/1.5 to 1.5 I have checked my data after feeding to model, there is sometimes huge object because they are very near to the camera.

I am curious how to solve this problem, if I don't use the multi-scale skill, the map of model will drop about 4 %, it seems that my model 'can't see' the object smaller than 44 pixels with 512512 size (about 20*20 pixels for source images 1080p)