This week, I presented our project proposal. I also spent time further researching different pre-trained ML model options and looking into different datasets available for use. As a part of this process, I took a closer look at the existing YOLO versions, and noticed that YOLOv5, YOLOv6, and YOLOv8 were not officially supported models unlike what I had previously assumed. I also discovered that YOLOv7 had a compressed YOLOv7-tiny version. Compared to YOLOv4, YOLOv7 has better mAP and flexibility on different compute architectures (evidence for two applications found here and here). Similarly, YOLOv7-tiny achieves the same AP as YOLOv4-tiny with less computational resources, making it further optimized for edge devices (with comparison evidence found here). I also looked into strategies for training YOLOv7 on custom data and found a good Google Colab setup reference here along with deployment and inferencing strategies here, specifically in regards to running trtexec for optimized model conversion and benchmarking. I’ve also found sufficient datasets online for plastic water bottles, crumpled paper, soda cans, and general recyclables.
Like my other teammates, I’ve also spent time bouncing back ideas for our pick-up mechanisms, verifying design sketch, further deciding parts, and updating our Gantt chart.
My progress is on schedule according to the chart, and for the next week I plan on successfully running YOLOv7-tiny and YOLOv7 on Google Colab, and also hopefully start training the model/set up custom data. Like my teammates, I’ll also be working on the design report and design presentation.
0 Comments