This is the 5th week of my journey, and it’s hard to believe that I am already halfway through the program! Despite feeling a bit unproductive this week, I made some progress in my research.

During this week, I focused on learning about a package called “rawpy.” It enables the loading of RAW images, working with their data, and creating postprocessed (demosaiced) images. This knowledge will be valuable for my future work in computer vision.

Computer vision is an extensive field with numerous subtopics to explore. I have decided to delve deeper into object recognition, specifically within the context of autonomous driving. One task that caught my interest is “vehicle detection.” To accomplish this task, I discovered the UA-DETRAC (Detection and Tracking in the Wild: Evaluation and Benchmark) dataset. It is a large-scale dataset designed for evaluating object detection and tracking algorithms in real-world traffic surveillance scenarios. However, I encountered a roadblock as the pre-trained models provided with the dataset were inaccessible for further fine-tuning.

To continue my research, I switched to another dataset available on Kaggle. I found a dataset that focuses on car object detection and fine-tuning using PyTorch here. Unfortunately, while attempting to replicate the work done by another researcher on this dataset, I encountered several errors related to environment setup. These issues undermined my confidence in relying on their work for further research.

As a result, I decided to change my focus from autonomous driving to a different task called “CatsOrDog.” This task involves image classification to distinguish between cats and dogs. I found a suitable dataset, code, and pretrained model for this task here. The codebase and the CNN model used in this task are relatively easier to understand, which will help me gain a better grasp of the concepts.

Fortunately, I connected with another researcher at our school who is also working on projects with Dr. Maxwell. She graciously shared her previous code on a similar project, which has proven to be a valuable resource for me. Their insights and code will contribute significantly to my progress.

Looking ahead to the next week, my plan is to recreate the “CatsOrDog” task. Additionally, I aim to explore the conversion of data from sRGB (jpg files) to raw data, which is close to linear sRGB data. I will be using png files for the conversion and log data with exr files.