Month: March 2025

For this week, I continued with refining the training pipeline for the object detection model, in order to maximize our performance for the demo . I expanded our dataset with the Open Images database, filtering out food-related classes to our use case, significantly increasing our existing data. Furthermore, I am looking to experiment with the YOLOv10 model in order to further increase accuracy.

I am currently on track with our milestones. Preliminary training have been completed, and we are working to maximize our accuracy and finalize our model. I am also working on deploying our model onto the Raspberry Pi.

For next week, we aim to proceed with our demo, during which I aim to show our model with the optimal accuracy. I will continue working on integrating our model with the tentative pipeline for fridge item detection, and continue expanding our training dataset and tuning our parameters to optimize accuracy.

1. Personal accomplishments for the week

1.1 Data transmission pipeline

This week, I continued debugging the data transmission pipeline on the Raspberry Pi that uses the boto3 Python package. The pipeline is able to reliably send files over, however, there appear to be latency issues with the time taken for transmission. I will have to investigate resizing the images before transmission.

2. Progress status

I am slightly behind schedule I have not started work on the motorized camera system construction. The data transmission pipeline developed, while functional, still presents some reliability issues.

3. Goals for the upcoming week

• Begin construction of motorized camera slider

For this week, I continued with developing my core image processing pipeline. I conducted tests on object classification using the YOLOv5x model with some sample fridge images, and I collaborated with my team to finalize the list of our target food items for model training.  I always begun with integrating the CV pipeline with the Raspberry Pi to validate real-time image capture.

For my progress, I am currently on track with my project timeline. Preliminary object detection is functioning, and I continuing to work on improving the model frequency.  I am beginning on integrating my model with the edge processor.

For my goals for next week, I aim to continue improving the performance of my YOLOv5 model, as well as integrating the model with our Raspberry Pi.

For this week, I continued with data collection, expanding our datasets by importing annotated images which I have found online. I have also performed further fine-tuning of our training hyperparameters, testing different augmentations and hyperparameter optimizations in order to improve our detection accuracy, incorporating the expanded dataset in order to improve the training process. I then ran a preliminary test for our YOLOv5x model to test its detection accuracy. As expected, the YOLOv5x model performs at a higher accuracy, though it results in a longer inference time when run locally.

In term of progress, I am on track by improving model accuracy while also maintaining real-time performance.  The next step in our progress chart is integration, and I will begin work with integrating our model with the Raspberry Pi and the model app.

For next week, I aim to run inference on cloud and measure inference timing in order to quantitatively assess the latency of our model. I will also continue to work on training and enhancing our YOLOv5x model, and conduct more real-world tests to measure the performance of our optimized model.