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

Object is never detected, random number of rectangles at random position #415

Open abhishekkumardwivedi opened 7 years ago

abhishekkumardwivedi commented 7 years ago
$ flow --model cfg/tiny-yolo-sample.cfg \
   --load bin/tiny-yolo-voc.weights \
   --train \
   --annotation ./sampleTraining/ann/ \
   --dataset ./sampleTraining/img/ 
$ flow --model cfg/tiny-yolo-lpc.cfg \
   --load bin/tiny-yolo-voc.weights \
   --imgdir ../sampleTest/img/

To confirm again, I used same sample images for training as well as testing. But on each test execution, boxes are at random location and keep changing on each execution. And also number of rectangle hilighters are also varying randomly from 3-4 rectangles to large number of rectangles. To again verify I tried testing on one single black image but test sometime highlights rectangle at random location on this black image file.

rlan commented 7 years ago

@abhishekkumardwivedi How is the loss value at the end of training? During test, did you try with --json and check the confidence value?

abhishekkumardwivedi commented 7 years ago

@rlan Below is the training complete screen, it shows avg loss of ~2.5. I didn't try for confidence value.

training-complete-screen

I did also post a stackoverflow query to debug the issue in slower system, without GPU but still looking for answer.