Accomplishment: For this week, I have accomplished training the object recognition model with an indoor object dataset I have found on the Internet. I have created a Google Collab, so that my teammates can also implement or view the progress of the training model and download the trained model after running all the steps. The model has changed from Yolov4 to Yolov5 because of the dependency issue in Yolov4. It uses darknet, which is now outdated and not maintained by the development team, so I have switched to Yolov5, which is developed by Ultralytics. This version is supported by a team that has a great maintainability, so this upgrade may be a huge advantage to our product.
However, it also means that I have to integrate a distance estimation feature from Yolov4 to Yolov5 model, which may take some time for development.
I have also collected a dataset relevant to indoor objects from the Internet. This dataset has 558 images for training, 52 images for validation, and 30 images for testing.
The challenge towards training the OR model is that the dataset needs to be annotated as well, so it might take more time to collect greater number of datasets. My plan is to search more data online or perhaps use some tools for annotating the images. I have found that Kaggle has some useful datasets, so I will take a look into the website for more data.
Progress: I am on schedule because I have already collected some indoor object dataset and trained the model with it. However, I am planning on collecting more dataset to increase accuracy and need time to add a distance estimation feature to Yolov5 model. Before spring break ends, I am planning on finish training the OR model with distance estimation feature.
Projected Deliverables: By next week, I will collect around 500 more images with annotations and train the model with it. I will also begin implementing distance estimation feature to Yolov5 model by cross-referring the original source of Yolov4 + distance estimation feature.