Training KITTI with DetectNet
YG
I have been playing with DetectNet over the last few days, and I have a few questions. I am using the KITTI dataset as a training example. The label format you are showing in your DetectNet blog has only 6 parameters (class, bounding box, and coverage). However, the label format in KITTI has many more parameters, but does not specify coverage. Yet, it works well on DetectNet. What is the right format then? Can I use the shorter format you are describing in your blog, or the KITTI format with default values for unspecified parameters should be used? Also, since the coverage parameter is not specified in KITTI format, how the network optimizes w.r.t. this parameter?
My second question is about the training itself. I am getting worse results than the ones reported in your GitHub page using the parameters and configuration you recommended. You reported a precision and recall of around 0.9 and mAP of around 0.7, where I am getting precision ~0.8, recall ~0.7 and mAP ~0.6. Can you explain these differences? I noticed that when I prepare the dataset according to your instruction, the splitting to training/validation is a bit different than in your example. I am getting a 6373/1108 split (train/val), whereas you have 5984/1496. Could it be the explanation for the differences in accuracy results?
