Click here to download the pdf of my Master's dissertation when I was studying in the University of Edinburgh in August 2019. The topic is "Understanding and affecting the accuracy-speed trade-off of object detectors for self-driving cars". I am proud of the research I did for CV based self-driving.
Experiments were done based on yolov3.
How can I use the weights file on my own computer?
predict a single image
./darknet detector test cfg/bdd_mini.data cfg/yolov3-bdd-final.cfg backup/yolov3-bdd-final_34000.weights input.jpg
Download the nNecessary files:
How was it trained?
Commands for the baseline model:
./darknet detector train cfg/bdd_mini.data cfg/yolov3-bdd-mini.cfg darknet53.conv.74 -map > baseline.txt 2>&1 &
./darknet detector map cfg/bdd_mini.data cfg/yolov3-bdd-mini.cfg backup_bdd_baseline/yolov3-bdd-mini_best.weights
./darknet detector valid cfg/bdd_mini.data cfg/yolov3-bdd-mini.cfg backup_bdd_baseline/yolov3-bdd-mini_best.weights
./darknet detector test cfg/bdd_mini.data cfg/yolov3-bdd-mini.cfg backup_bdd_baseline/yolov3-bdd-mini_best.weights
Some key configurations for my model:
- resolution: 896x512
- complete + balanced training set
- 5 scales
- 13 anchors
- using softplus
I am sorry that I can not remember the accurate steps. I can only find this training log I recorded for the final model:
15000
0.231 0.490 0.190 0.097 0.273 0.378
continue at 611526
16000
0.231 0.489 0.188 0.102 0.272 0.386
24000
0.241 0.505 0.199 0.109 0.287 0.390
continue at 617163
lrx0.1 at 25000
30000
0.270 0.542 0.233 0.121 0.316 0.442
lrx0.1
32000
0.274 0.545 0.239 0.121 0.318 0.452
continue from 32000 with bdd_train_balanced_4
lrx0.1
34000
0.277 0.549 0.244 0.121 0.322 0.469