search

thtrieu / darkflow

Translate darknet to tensorflow. Load trained weights, retrain/fine-tune using tensorflow, export constant graph def to mobile devices
GNU General Public License v3.0
6.13k stars 2.08k forks source link

Boxes appear on the left side only. #1049

Open GioLomia opened 5 years ago

GioLomia commented 5 years ago

Trained on my own data set. 8 classes, at step 11790 - loss 0.15480981767177582 and going down, I changed the number of classes in the config as well as the filters on the last layer to 65 (5*(classes+5)). Here is the command I used for training : !python3 flow --model cfg/yolov2-voc-8c.cfg --train --dataset ../Training_Set/Images --annotation ../Training_Set/Annotations --trainer adam --batch 10 --save 100 --keep 1 --gpu 1 --summary tensor --epoch 50 --load -1

!python3 flow --imgdir ../Training_Set/Images --model cfg/yolov2-voc-8c.cfg --load -1

I am having issues with bounding boxes. They do not appear unless I lower the threshold to some ridiculously low value like 0.008. And when they do appear they are tiny boxes and they all form a straight line on the left border of the picture.

kmarrip commented 5 years ago

check the annotations file, check if the bounding boxes are exactly where u want them to be.

rockstardotb commented 5 years ago

@GioLomia When you first start training on custom data, you should overfit the network on a subset of images (3 - 5). keep training until you get perfect bounding boxes on those images with 0.9 confidence or greater. After that, begin training on the entire dataset.

I had a similar issue and this fixed it for me.