Giters

BichenWuUCB / squeezeDet

A tensorflow implementation for SqueezeDet, a convolutional neural network for object detection.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

About setting Anchors on Custom Datasets

tejashadkar opened this issue · comments

commented

Dear @BichenWuUCB ,
Thanks for the code and explanation on how to use it! I am trying to train a restructured model on custom dataset. In #53 you mentioned about possible reasons for top matched anchor only has an IOU of 0.1. Probably, if my code falls in one of the category where the number of center points and anchors are not enough for matching ground truth. I have two questions:

  1. I tried setting number of anchors to 30 on experimental basis, still I have IOU's max reaching 0.7 and lot of objects missing IOU's. I tried increasing number of center points and anchors such that spatially they will be placed at every alternate pixel. IOU shoots up to 0.9 but there are always some objects with IOU=0. Any suggestion how do I fix that?

  2. Do I have to always change the size of last ConvDet layer according to the size of the anchors? Is there a different way to do this without increasing last layer size? I am using GTX1060 6gb and model does not fit in to this size of memory. Any suggestions?

IOU:

max iou: 0.7383245945616387
min iou: 0.0
avg iou: 0.09621441173926668
number of objects: 26.0
number of objects with 0 iou: 11
Warning: Discarded 0/(26) labels that are assigned to the same anchor
max iou: 0.34908415184496944
min iou: 0.0
avg iou: 0.04020778139716306
number of objects: 25.0
number of objects with 0 iou: 14
Warning: Discarded 0/(25) labels that are assigned to the same anchor

Please let me know, any suggestion would be helpful!

commented

I found mistake, as I was referring to image data as row major because of python and numpy arrays while all methods from source code were implemented as column major following image co-ordinate system.