We have a robot that roams around our house and we want to create an object detection model to help it recognize things it shouldn’t run in to, namely, table legs and chair legs. Some of our chair legs are quite narrow and black, while chairs legs are 3x as wide and a natural brown wood color. We also have “squat” legs for some book cases and other furniture.
My question is, should we be training it to recognize 3 types of legs, or train it to recognize just general furniture legs? Any thoughts on which is the better approach?
1 Like
If you’re just looking for objects to avoid, my guess would be to train a model to simply look for “legs” and use a variety of furniture legs at various distances and lighting conditions to train the model. Transfer learning with MobileNetV2 SSD works pretty well for these types of applicaitons.
You can definitely try both approaches to see which works better. You would just need to create 2 projects in Edge Impulse and re-label the data with the 3 different types of legs to compare results to the generic “furniture leg” detection approach.
1 Like
Thanks, I’ll probably end up trying both approaches.
1 Like