search

amazon-science / progressive-coordinate-transforms

Progressive Coordinate Transforms for Monocular 3D Object Detection, NeurIPS 2021
Apache License 2.0
64 stars 10 forks source link

How to generate each learner's label? #4

Open rockywind opened 3 years ago

rockywind commented 3 years ago

How to construct the location loss function? image

Willy0919 commented 3 years ago

How to construct the location loss function?

The target for each stage is the ground truth localization, the loss function calculates the loss between ground truth localization and predicted localization of the corresponding stage. For stage t, we regress the center localization residual based on the previous stage prediction, so the predicted localization is the sum of the initial center localization and a stack of residual localization from t stages, which can be described in Equation 4 in our paper. Besides, the implementation can be seen in lib/models/patchnet_pct.py and lib/losses/patchnet_loss.py.