search

rbgirshick / py-faster-rcnn

Faster R-CNN (Python implementation) -- see https://github.com/ShaoqingRen/faster_rcnn for the official MATLAB version
Other
8.11k stars 4.11k forks source link

Re-trained model gives better Median AP than base model but the confidence levels are REALLY low. #360

Open GeorgiAngelov opened 8 years ago

GeorgiAngelov commented 8 years ago

So I am using the original /data/faster_rcnn_models/py-faster-rcnn caffe/VGG16_faster_rcnn_final.caffemodel caffe model for my weights and I am re-training it. My data classes are a subset of the original pascal_voc VGG16 classes so I left all files for the model intact. What I did is I just started further training the model with my own data.

I ran the test_net.py script and I got the following results:

Original model: Mean AP = 0.0237 My model after 20k iterations: Mean AP = 0.5659

Judging by this, my model should be significantly better than the original model. However, when I actually get the detection from my model, my confidence is extremely low. Wouldn't this indicate that the test_net results contradict the real-world example?

BadWindshield commented 8 years ago

Are you sure the original precision score is ~0.02? Precision is the probability of detecting something given that we "know" it's there.

GeorgiAngelov commented 8 years ago

@BadWindshield, where would I have changed this value? I am not sure I've encountered it so far.