# Sampling + Fitting

Inputs: Features + Sampling Weights

1. Sample bounding box based on sampling weights
2. Encode features in BB with ResNet
3. Predict partial occupancy function based on feature encoding
4. Project depth map in BB into 3D-Space
1. Calculate projective transformation for coordinate system
2. Normalize to unit cube
5. Combine partial occupancy function with previous partial functions
1. For each partial occupancy function store the corresponding transformation and apply it to the input query before averaging the occupancy result with Gaussian averaging
6. Sample N points in bounding box in input features (maybe all pixels?) and determine inlier score based on sum of distances to mesh
1. What about sampling in the 3D space?
2. How can this be made occlusion aware?
1. Sampling on a ray farther away and closer to the camera and determining the distance to mesh

# State
* Maximum inlier score
* For each size of bounding box?