search

philipperemy / yolo-9000

YOLO9000: Better, Faster, Stronger - Real-Time Object Detection. 9000 classes!
Apache License 2.0
1.18k stars 309 forks source link

Why last layer have dimension 28269? #16

Closed DC-Shi closed 6 years ago

DC-Shi commented 6 years ago

I want to fine-tune my own image-label based on yolo9000 weights. Unlike yolo weights, I can compute last layer filter = (n+1)*5 for n labels.

But 28269 filters gives 9418 features, not something like (9418+1)3. I only find 28269 = 94183 + 15.

For the YOLO9000 paper, it just showed darknet-19, it's not end with 28269 filters(but 771000).

DC-Shi commented 6 years ago

Oh, finally I found some explanation:

In YOLO9000, it gives 3 anchors(yolo has 5 anchors), so final filters should be multiple of 3. For each anchor, it must have: Probability of the detected object, x,y, width, height (5 numbers) And with 9418 labels we want, we finally get (9418 + 5)*3 = 28269