Ultra-Fast-Lane-Detection
Ultra-Fast-Lane-Detection copied to clipboard
cfzd
how to achieve the acc in the paper
I had trained the network with backbone is res18,but cannot attach 68.4 f1-score in culane dataset. Coudle you help me? thanks
@FANG-MING Are you using the default training parameters? If you are using the default training parameters, just run
sh launch_training.sh
Then you can get the results. The provided models are trained in exactly the same way.
@FANG-MING If you want to train with 1 GPU, you should reduce the learning rate from 0.1 to 0.025.
@FANG-MING If you want to train with 1 GPU, you should reduce the learning rate from 0.1 to 0.025.
how can i test the speed? I find I can attach the spend descibed in the paper.
@FANG-MING What's your device? What're the results of running speed_simple.py?
@FANG-MING What's your device? What're the results of running speed_simple.py?
I test in 1080ti, but got speed is 260fps fastest
@FANG-MING it's possible since the latency is already so low. The inference time for 300fps is 3.3ms. For 260fps, it is 3.84ms. Any latency fluctuation could make the speed slower.
Hello, author, why can't I reach the accuracy of the paper when I reproduce it with the official code of resnet34 network
@ma-cg It is possible since the released repo is a re-implementation. We will release a new version very soon, in which the res34 could achieve a ~76 F1 measure.