max-unfinity / clickseg-cracks

Training ClickSEG HRNet18 on custom Cracks Dataset

Origingal repo: https://github.com/XavierCHEN34/ClickSEG

---

Generally, it takes only 1-3 clicks avg. to achieve good cracks predictions, but the model often do the job in 1 click.

Trained checkpoints: https://drive.google.com/drive/folders/1v7KxJenZzYii6Dh7Ad2ePKn4SinSS4Q7?usp=sharing

Prediction Samples:

Dataset Info

Crack-segmentation-dataset from kaggle: https://www.kaggle.com/datasets/lakshaymiddha/crack-segmentation-dataset

This Dataset contains around 11.200 images that are merged from 12 available crack segmentation datasets.

There are many inaccurate masks in ground truth. This adds noise to the data, reducing the IoU metrics significantly.

Inaccurate masks samples:

Trained model info:

• Model: HRNetV2-W18 S2
• Pretrained: ImageNet
• Input size: 256, 256
• Train data count: 9603
• Batch size: 40
• Epochs trained: 23
• Val. Metric (AdaptiveIoU): 0.5788

Val. Metrics (FocalClick hrnet18), 500 test samples:

[email protected]%	[email protected]%	[email protected]%	[email protected]%	[email protected]%
1.76	2.29	2.91	3.71	4.87

IoU at 1 clicks	IoU at 2 clicks	IoU at 3 clicks	IoU at 5 clicks	IoU at 15 clicks
0.6280	0.6340	0.6386	0.6504	0.6825

There are several changes from the original clickSEG repo:

• Fixed data loading in ISDataset. Without this fix some training samples was skiped (isegm/data/base.py)
• Added validation step with NoC metrics while training
• More augmentations in training
• Updated to last version of albumentations
• Removed dependency on mmcv (for HRNet)
• Better eval script, more metrics