Bento Collares Goncalves

Muslim prayer, or Salah, involves repeating specific cycles of movements and prayers, known as Rakats. Tracking these can be crucial for people with memory impairments or distractions. To help with this, we've developed a computer vision model that can accurately count these prayer cycles.

Using a small dataset with 21 complete Rakat cycles and the correspondent cycle count for a video as annotation, I developed a lightweight computer vision approach that can accurately capture key transitions within Rakat cycles to keep track of completed Rakats.

The approach combines intensity thresholds, pauses, and comparison with reference images to detect transitions into prostration during prayer. To calibrate model parameters, I employed a Bayesian hyperparameter search that converged on a solution that performed well across all labeled examples.

This model, now available on App Store, has the potential to be a valuable tool for assisting individuals, especially those with disabilities, in accurately performing and completing their prayers.

Pack-ice seals in the Southern Ocean are good indicators of environmental conditions. Their size and widespread distribution make them perfect for monitoring through high-resolution satellite images. However, it's challenging to manage the massive amount of images by manually identifying the seals.

SealNet 2.0 is an automated system that can detect seals. It uses one model to find potential seal habitats by identifying sea ice and several more models to find the seals themselves.

The system achieves a precision of 0.806 at 0.64 recall in a robust, undisclosed test set, outperforming two human experts and the older version of SealNet. It achieves this improvement by focusing on images of sea ice only, fine-tuning its settings with the help of high-performance computing, and refining predictions based on statistical analysis.

Even a simplified version of this system can improve the accuracy of seal detection by human experts. It could also help train new experts. However, like humans, the system struggles with rugged terrain. So, we must use statistical methods to adjust the seal population estimates it produces.

Antarctic penguins are vital for understanding the health of our environment, particularly in the face of climate change. We've developed a Deep Learning model to identify Adélie penguin colonies from high-resolution satellite images.

To teach our model how to identify penguin colonies, we used the Penguin Colony Dataset, which includes over 2,000 images from 193 colonies. Due to a lack of detailed labeling of these images, we've developed a method to learn effectively from less precise labels.

We used a system that could sort out data unsuitable for this learning process. The learning process is trained using a specific calculation that can learn effectively from less precise labels. Our tests have shown that this less-precise labeling can significantly improve the model's performance. The model's accuracy in identifying penguin colonies increased significantly when we included these less precise labels in its training, improving IoU from 42.3% to 60.0% at a held-out test set.