IndexError: list index out of range

[fayechou]

And I print the value: print (feature_list_new):

Could you help to find out the problem? Thx~

[cardwing]

@fayechou, since you use a different input size (256 x 512 instead of 288 x 800), then the parameters of the message passing module should also be changed. Since message passing is performed in rows and columns, then you need to modify the index number from 35 to 31 in row dimension and 99 to 63 in column dimension. Just replace 35 to 31 and 99 to 63 should help. Besides, remember to modify this line ret['prob_output'] = tf.image.resize_images(conv_output, [288, 800]), the output size should be 256 x 512. I will remove the hard code part soon.

[fayechou]

@cardwing Thanks for your detailed reply. I am able to test the CULane dataset now with input size 288*800. And I am training the TUsimple data according to your above instructions.
One question: does this network can only detect up to four lane markings? Thanks! I know in TUsimple dataset groundtruth, some frames have 5 lane markings.

[cardwing]

@fayechou, you can just modify the output of the model. More specifically, modify this line conv_output = self.conv2d(inputdata=dropout_output, out_channel=5, kernel_size=1, use_bias=True, name='conv_6') and this line fc_output = self.fullyconnect(relu_output, 4) in vgg_encoder.py, the output channel should be 6 (5 + 1) and the output of the fc layer should be 5. Modify other files if it raises errors related to the output dimensions.