TuSimple results

LANEDETECTION.md

@@ -39,12 +39,24 @@ Mixed precision training on CULane with ERFNet:
 python main_landec.py --epochs=12 --lr=0.15 --batch-size=20 --dataset=culane --model=erfnet --mixed-precision --exp-name=<whatever you like>
 ```
 
+Mixed precision training on TuSimple with ERFNet:
+
+```
+python main_landec.py --epochs=12 --lr=0.7 --batch-size=20 --dataset=tusimple --model=erfnet --mixed-precision --exp-name=<whatever you like>
+```
+
 Mixed precision training on CULane with ERFNet-SCNN:
 
 ```
 python main_landec.py --epochs=12 --lr=0.08 --batch-size=20 --dataset=culane --model=scnn --mixed-precision --exp-name=<whatever you like>
 ```
 
+Mixed precision training on TuSimple with ERFNet-SCNN:
+
+```
+python main_landec.py --epochs=12 --lr=0.3 --batch-size=20 --dataset=tusimple --model=scnn --mixed-precision --exp-name=<whatever you like>
+```
+
 ## Testing:
 
 Training contains online fast validations by using --val-num-steps=\<some number > 0\> and the best model is saved, but we find that the best checkpoint is usually the last, so probably no need for validations. For log details you can checkout tensorboard.

README.md

@@ -64,8 +64,8 @@ And models from this repo is faster (also better or at least similar) than the o
 | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
 | ERFNet | 288 x 800 | *yes* | CULane | 0.7346 | 0.7359 | 6h |
 | SCNN | 288 x 800 | *yes* | CULane | 0.7392 | 0.7405 | 11.3h |
-| ERFNet | 360 x 640 | *yes* | TuSimple |  |  |
-| SCNN | 360 x 640 | *yes* | TuSimple |  |  |
+| ERFNet | 360 x 640 | *yes* | TuSimple | 0.9468 | 0.9483 | 0.2h |
+| SCNN | 360 x 640 | *yes* | TuSimple | 0.9520 | 0.9526 | 0.4h |
 
 
 *\* All performance is measured with ImageNet pre-training and reported as 3 times average/best on test set.*

code/all_utils_landec.py

@@ -17,7 +17,7 @@
 def erfnet_tusimple(num_classes, scnn=False, pretrained_weights='erfnet_encoder_pretrained.pth.tar'):
     # Define ERFNet for TuSimple (With only ImageNet pretraining)
     return erfnet_resnet(pretrained_weights=pretrained_weights, num_classes=num_classes, aux=4,
-                         dropout_1=0.1, dropout_2=0.1, flattened_size=4400, scnn=scnn)
+                         dropout_1=0.3, dropout_2=0.3, flattened_size=4400, scnn=scnn)
 
 
 def erfnet_culane(num_classes, scnn=False, pretrained_weights='erfnet_encoder_pretrained.pth.tar'):
@@ -42,7 +42,7 @@ def init(batch_size, state, input_sizes, dataset, mean, std):
          Normalize(mean=mean, std=std)])
     transforms_train = Compose(
         [Resize(size_image=input_sizes[0], size_label=input_sizes[0]),
-         RandomRotation(degrees=1),
+         RandomRotation(degrees=3),
          ToTensor(),
          Normalize(mean=mean, std=std)])