search

matterport / Mask_RCNN

Mask R-CNN for object detection and instance segmentation on Keras and TensorFlow
Other
24.54k stars 11.68k forks source link

training dataset size for transfer learning is 100 to 1000? #1531

Open nyck33 opened 5 years ago

nyck33 commented 5 years ago

When I ran mask-r-cnn on the tennis match video as is, it did not even detect the tennis ball at all. Just the players and crowd mainly which was weird because there is a sports ball class.

Still, I want to see if mask-r-cnn can track a tennis ball if I use transfer learning and set the config to just 2 classes: background plus this new "ball" from the annotated image frames below.

I am annnotating frames of a tennis broadcast per below:

image

image

image

Note; this used Annotator version 2 but I have switched to version 1 as ver.2 is not compatible with the code as is.

nyck33 commented 5 years ago

I was able to make some tweeks to the code to use the circular bounding shape but it looks inaccurate with only a training set size of 150 images and validation of 50 images.

How many do I need of each to get the accuracy up?

As you can see it thinks the ball is near the bottom left corner off the court:

image

ipynb notebook and adjusted balloon.py are in my repo: repo