Topic: Modification of YOLOv3(Keras) for object distance estimation
(original model: https://github.com/qqwweee/keras-yolo3)
Methods:
1.Modify the output of YOLOv3 by adding a predicted distance variable at each prediction layer.
2.Meanwhile, modify the original YOLOv3 loss function by adding a distance loss function (here I tried MAE,MSE,RMSE. MAE is the best among them).
3.Retrain the modified model by main using processed public KITTI dataset and inhouse SMART dataset. (Transfer Learning, Finetune...)
4.Evaluation: judge the performance of object detection (metrics of TP, PR curve, mAP) and distance estimation separately.
Some test examples are shown here:
