A few train-related questions about tiny

[peiyunh]

1. Given a linear predictor, concatenating features is equivalent to predicting based on individual feature and summing them up.
2. I would run the detector with an image pyramid but without 2X upsampling. If you give up 2X, it should only hurt the performance on tiny-sized faces.
3. That is strange. What I saw when I add hard negative mining into our detector is that results look qualitatively much nicer because of less false positives, but it does not yield much improvement in the final accuracy.
4. This is what I learned from how FCN is trained. I believe they use x100 smaller learning rate when training FCN8s at once but I found x10 works fine.
5. I believe there are papers initialize predictor with zero weights, such as FCN.

[thealchemist-x]

Hi @peiyunh

Thanks for your amazing work. I'm new to deep learning and am interested to pick up small faces. If I may ask a simple question on the training dataset:

Q) If I have generally large faces in my datasets (example, a face occupying 120 pixels x 70 pixels from a larger image resolution of 640 x 480), do I have to crop the face out and resize to say 30 pixels x 30 pixels for the region of interest and crop the face together with the context and resize to say 60 pixels x 60 pixels before training?

Thank you for your guidance!

[peiyunh]

Your idea makes sense and should fit better into a two-stage architecture, such as Faster RCNN.