search

sacmehta / ESPNet

ESPNet: Efficient Spatial Pyramid of Dilated Convolutions for Semantic Segmentation
https://sacmehta.github.io/ESPNet/
MIT License
541 stars 112 forks source link

About IoU on Cityscapes. #34

Closed wangq95 closed 5 years ago

wangq95 commented 5 years ago

Hi, I was confused how to achieve the performance reported on Cityscapes' benchmark with IoU of 60.3%? How to set the modelType and the 'p' and 'q' in VisualizeResults.py? Thanks!!

sacmehta commented 5 years ago

See #8 #31

wangq95 commented 5 years ago

I am sorry that this issues are not mentioned how to choose the 'modelType' and the 'p' and 'q', can you tell me directly? Thank you!

sacmehta commented 5 years ago

parser.add_argument('--modelType', type=int, default=1, help='1=ESPNet, 2=ESPNet-C') parser.add_argument('--decoder', type=bool, default=True, help='True if ESPNet. False for ESPNet-C') # False for encoder parser.add_argument('--weightsDir', default='../pretrained/', help='Pretrained weights directory.') parser.add_argument('--p', default=2, type=int, help='depth multiplier. Supported only 2') parser.add_argument('--q', default=8, type=int, help='depth multiplier. Supported only 3, 5, 8')

These are the arguments that you need to change in Visualize*.py file. Could you please be more specific what is not clear about these arguments?

wangq95 commented 5 years ago

Nope, I have understood. Thanks a lot!

wangq95 commented 5 years ago

I got a new question. When I use the model ‘espnet_p_2_q_8.pth’ to predict on TEST set of Cityscapes(1525 images), remap and submit the predicted result to Cityscapes benchmark, and only got IoU of '57.2178', but the reported value is 60.3. I was confused.

sacmehta commented 5 years ago

See #8 #9

uvinodkumar commented 5 months ago

I got a new question. When I use the model ‘espnet_p_2_q_8.pth’ to predict on TEST set of Cityscapes(1525 images), remap and submit the predicted result to Cityscapes benchmark, and only got IoU of '57.2178', but the reported value is 60.3. I was confused.

Hello I haven't trained the models i want to use the same provided model ‘espnet_p_2_q_8.pth’ in the github and want to test on cityscapes test data and calculate IoU and check results for benchmarking. Could you please explain the process for getting IoU for cityscapes test images.