about IoU with different size from cityscape,please help me!

[electronicYH]

Because of the camera output size is 640*360, so I changed the size of cityscapes dataset. Then I use the project code to train, but I can't get the good IoU about 75%, please help me , how can get the good IoU result.
the train log is:
INFO:ptsemseg:Iter [90000/90000] Loss: 0.9804 Time/Image: 0.0165 lr=0.090953
11it [00:05, 2.08it/s]
INFO:ptsemseg:Iter 90000 Val Loss: 1.1422
Overall Acc: 0.902127139034985
INFO:ptsemseg:Overall Acc: : 0.902127139034985
Mean Acc : 0.6061411175875274
INFO:ptsemseg:Mean Acc : : 0.6061411175875274
FreqW Acc : 0.8350029456995808
INFO:ptsemseg:FreqW Acc : : 0.8350029456995808
Mean IoU : 0.48783643289422235
INFO:ptsemseg:Mean IoU : : 0.48783643289422235
INFO:ptsemseg:0: 0.949375256023792
INFO:ptsemseg:1: 0.6434965931893878
INFO:ptsemseg:2: 0.8355899486279862
INFO:ptsemseg:3: 0.36742551411680824
INFO:ptsemseg:4: 0.24889070738206942
INFO:ptsemseg:5: 0.31642120260993717
INFO:ptsemseg:6: 0.28094440896774303
INFO:ptsemseg:7: 0.42495380920802917
INFO:ptsemseg:8: 0.8470327817169933
INFO:ptsemseg:9: 0.4534516073428247
INFO:ptsemseg:10: 0.8743228190634068
INFO:ptsemseg:11: 0.4587472314322872
INFO:ptsemseg:12: 0.310287655352762
INFO:ptsemseg:13: 0.8298468752303512
INFO:ptsemseg:14: 0.42321140100466925
INFO:ptsemseg:15: 0.39785631228298224
INFO:ptsemseg:16: 0.031675380114154154
INFO:ptsemseg:17: 0.09916433094570747
INFO:ptsemseg:18: 0.4761983903783339

and the hardnet.yml is:
model:
    arch: hardnet
data:
    dataset: cityscapes
    train_split: train
    val_split: val
    img_rows: 360
    img_cols: 640
    path: ../cityscape_transformation/
    sbd_path: ../cityscape_transformation/
training:
    train_iters: 90000
    batch_size: 48
    val_interval: 500
    n_workers: 8
    print_interval: 10
    augmentations:
        hflip: 0.5
        rscale_crop: [360, 360]
    optimizer:
        name: 'sgd'
        lr: 0.1
        weight_decay: 0.0005
        momentum: 0.9
    loss:
        name: 'bootstrapped_cross_entropy'
        min_K: 4096
        loss_th: 0.3
        size_average: True
    lr_schedule: 
        name: 'poly_lr'
        max_iter: 9000000
    resume: None
    finetune: None    
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
and I only modify the code train.py here:
    v_loader = data_loader(
        data_path,
        is_transform=True,
        split=cfg["data"]["val_split"],
        img_size=(360,640),
    )
and the code cityscapes_loader.py here:
    def __init__(
        self,
        root,
        split="train",
        is_transform=False,
        img_size=(360, 640),
        augmentations=None,
        img_norm=True,
        version="cityscapes",
        test_mode=False,
    ):
        """__init__

[PingoLH]

Hi, honestly, you can't get a good IoU for such a small input resolution, since this network architecture was designed for 2048x1024 input. You can notice that the first four conv layers have two layers with stride=2, which means the spacial information was quickly shrunk into 512x256, while in your case, it will be 160x90 which is way too small for segmentation. You can try to simply remove the "stride=2" for both or one of the mentioned two conv layers (line #270, #272). The pretrained weight can still be loaded, but the network inference speed will be much slower.

[electronicYH]

Thank you for your help, I will try it. Thank you!

[MrCrazyCrab]

@electronicYH have you solve the problem to small input image size?