Water Segmentation Models #62
Comments
|
Experiment 1 : Training on puddles & Infering on flooded scenes
Remark : We trained the network with the same parameters as given by the author of the paper except for the epoch number (50K vs 60K). Infering on flooded images while training on puddle dataset doesn't work well. See qualitative results : puddle_detection failed_case_flood_segmenter Conclusion : I suggest we wait for the author to answer and deliver the supplementary materials where there is a detailed analysis on training time, robustness to over-fitting etc.. If we agree to spend more time testing this method, we should train it over our own flooded scene. |
|
Here is a Link to the fork. |
|
Experiment 2 : Initializing with the puddle model and training over 50 flooded segmented images
Remark 1 : This experiment aim to evaluate the ability of such a network to segment flood, results are promising for a fine-tuning. We could try to extract soft segmentation mask and combine it with other methods with fuzzy theory. Remark 2: The model hasn't been shown images without water so it is biased, one may want to train the latter with such images.
Conclusion : |
|
To do :
|
|
Update : Reading this github issue Starting A8F6 from the observation that GauGAN can draw pretty realistic water, I look deeper into their paper and github issue to find which was the semantic segmentation model used in their experiment with flickr landscape dataset. Following this github issue, they used DeepLab v2 on COCO-stuffs-164k . Evaluating Results A good way of using it would be to process our entire dataset and remove manually every image with less than 50% of water covered. If we do that post evaluation on the 210 images, we keep 88 % of them and increase the IOU to 86%. Good Results: Worst Results: The algorithm fails to recognize water, when it recognize the road, it also happens when the image is in heavy rain condition (few centimeters over the road and we still see the lane) |
|
Here is the implementation of the selected algorithm: https://github.com/cc-ai/deeplab-pytorch/
|
We are planning to work on segmentation models that have been designed specifically for the water scenario :
The approach used is described in Single Image Water Detection. The net is designed to detect water on asphalt and country road using reflection unit.
The paper code : Github Link
The text was updated successfully, but these errors were encountered: