Lanbig

My fun mini weekend project on Object Recognition. I’m very impressed on the evolution of Computer Vision. The computer started recognizing the objects by using ~200 images for the training data. The project was performed using an NVIDIA Tesla K80 and training for ~12 hours on Google Cloud Compute Engine.

How do I train the algorithm?

Step 1: manually label region (object/element) in each image (heavy lifiting work!).

Labels will store (x,y) coordinates of boundary box in the picture with class/type of the box.

Step 2: Feed labeled image to object detection algorithm and select pre-trained deep learning model.

I’ve chose “Faster R-CNN” model becasue it recieved a good score of mAP (mean Average Precision). However, it takes longer time to process images than other models.

Step 3: Monitor training process closely through an algorithm dashboard.

Step 4: Validate model by applying models on images that are in test set.

Check out GitHub https://github.com/Lanbig/custom-object-detection