Accomplishment:
This week, I have downloaded and tested several object recognition models that have distance estimation features to see how much implementation and add-on are needed to make a viable model. What I have realized about the two models I have first tested is that they both base off the object recognition model from the Yolov4 module. The techniques for ML are similar, and when going through the codes, I have found some space to adjust the model to be a viable model for our project.
During the testing, I have also realized that their pre-trained models detect a few objects (human, cellphone, etc.) that are irrelevant to indoor objects. Therefore, I am planning on finding a new dataset to train the model myself. Then, I have tested the Yolov7 model by installing NVIDIA CUDA and running the test to determine whether the model accurately detects a specific object that has been used to train the model. By comparing the accuracy based on the research papers on Yolov4 and Yolov7, I determined that using the Yolov4 model with distance estimation is sufficient to become our object recognition model.
I have also updated the software flow diagram by adding more specific details on inputs, outputs, and features that go with the data processor.
Progress:
I am on track on testing object recognition models, but because I have added a few more objectives in terms of training the chosen model with a suitable indoor object dataset, the planned date of testing the finalized object recognition model has been postponed by a week.
Projected Deliverables:
By next Tuesday, I will be done searching for a dataset with common indoor objects. If time permits, I will include some partial images of respective objects to take consideration of identifying an object even if the object is too close. By February 26th, I will be done training the model with the dataset that I have collected on Tuesday.