Recruitment Project for 2nd Year
While using GPS mapping services like Google Maps, one might observe that they are able to mark out where roads or travelable paths are located on the map. The proposed problem is similar: Given an image taken by satellite, segment out road pixels from those that aren’t roads (for e.g., pixels that include buildings, trees, water bodies etc). The aim shall be to classify individual pixels to generate a black-and-white binary mask for the original image, depicting roads with white pixels and everything else in black. This problem is known as Semantic Segmentation in the domain of Computer Vision.
We used DeepLabV3+ model to segment out road pixels from non-road pixels.
The dataset that we used to train and validate the model is Massachusetts Roads Dataset.
Since the images were quite large (about 1500 pixels x 1500 pixels), to get better segmentation results, the dataset was preprocessed before training. We cut the images to a smaller size (256 pixels x 256 pixels, segmented out the roads for these new smaller images and finally stitched back the output masks that we got.
The architecture of DeepLabV3+ looks like:
The results produced were like:
The overall process can be summarised as:
- Shrashtika Singh
- Apurba Prasad Padhy
- Ayushi Raj
- Tushar Sahu