This work is about improving building instance segmentation through Deep Structured Active Contours DSAC on Bing Huts dataset using image augmentation techniques like Denoising and Super Resolution.
The work highlights how Denoising and Super Resolution can help in removing noise or increasing resolution respectively, thereby improving quality of images, can consequently lead to better segmentation.
The work is specifically done on Bing Huts dataset because of its low-resolution images of 64 pixels. Following results highlight the improvement in evaluation metrics with augmented datasets:
| Dataset | IoU |
|---|---|
| Bing Huts (64 X 64) | 0.64 |
| Denoised Bing Huts (64 X 64) | 0.66 |
| Bing Huts (128 X 128) | 0.68 |
From the results, we can conclude that our 'enhanced' data has marginally been able to improve building segmentation. Such enhancing techniques can be applied in general to low-quality images in order to feed the model with better input.
No modifications were required in the code for Denoised dataset. However, for higher-resolution images, the building footprints were needed to be modified to match building boundaries as shown below:
- Use
make_building_footprints.ipynbto transform building footprints and store them in a CSV file.
Also, in order to expedite the learning process for higher-resolution images, the initial ACM was initialized with double its original size by a radial outward transformation (similar to that done on building footprints). This helped in better initialization and made the learning faster.
- Refer to
main_bing_sr.pyto see modifications regarding ACM initialization.
A pretrained GAN-based MSRResNet was used to obtain denoised images. The average PSNR obtained between original and noisy images was 30.10 dB. The following is an image-pair highlighting the model performance:
A pretrained (Residual Dense Network) RDN was used to obtain 128 X 128 higher-resolution images.