Training model with only drivable area segmentation

[danielzhangau]

I am trying to train the model with only one task - drivable area segmentation. I have been trained on the machine with GTX 1080Ti for three days already. Now in 180 epochs, the total training epoch gonna be 240. I checked images in the runs/visualization folder. By monitoring the Tensorboard, the train_loss is slowly decreasing but seems like not going to have a big change. The real-time training result comparing gt and my model prediction is below: I feel like my model prediction at this stage have not reached my expectation. My question is: will training with only one task affect model accuracy on the single task? If yes, How can I just run the model with only the drivable area been detected? I am trying to make the model lighter with doing less tasks note: I have not changed other configs.

Any helps will be appreciated!!

[hetao828]

I also encounter this problem when I train my lane line segmentation, the trained detection head gets worse.

[danielzhangau]

I also encounter this problem when I train my lane line segmentation, the trained detection head gets worse.

Hi, I retrained my model with an encoder and single detection head to make the model performs better than doing three tasks. You can try to modify the train options.

[dinarkino]

I think the reason for such a behavior is that option _C.TRAIN.DRIVABLE_ONLY = True does not include encoder training.

[sparshgarg23]

@danielzhangau just wanted to follow up ,did you set the following options to true

_C.TRAIN.ENC_SEG_ONLY = True
_C.TRAIN.DRIVABLE_ONLY = True