Questions about keypoints

[ChengkaiYang]

Dear @cocoon2wong : What a great job in trajectory prediction! But I still have some questions. In AKL loss, we supervise coarse trajectory with keypoints. How do we get ground truth keypoints used for loss function? Are ground truth keypoints part of ground truth trajectory or ground truth keypoints are constructed artificially based on ground truth of trajectory? If keypoints are constructed artificially based on ground truth, how can we guarantee it will bring more efficient or extra information to model to learn? (Imagine if we don't use such method, ground truth may be the only thing to supervise.If anything irrelevant or not such useful brought in model to supervise, does the performance of model drop?) If keypoints are part of ground truth, my question will be solved. Looking forward for your reply! Thank you very much!