Currently under dev. :param x: image blob () :param upsampleH: New H obsolete :param upsampleW: New W obsolete

Currently under dev. :param x: image blob () :param return_feature: True to extract features, False only return pose predicti

use nerf render imgs instead of use real imgs

get relative pose from abs pose pose1 to abs pose pose2 R^{v}_{gt} = R_v * R_gt.T :param: pose1 [B, 3, 4] :param: pose2 [B, 3, 4] re

Check if current model the best one. get early stop counter, if counter==0: it means current model has the best validation loss

load posenet training data

sample from log uniform distribution between 0.01~1

compute loss only in masked region :param criterion: loss function :param f1: [3, batch_size, H, W] :param f2: [3, batch_size, H, W]

Find unit norm of channel wise feature :param: tensor, img tensor (C,H,W)

normalize xyz into [-1, 1], and recenter pose

normalize xyz into [-1, 1], and recenter pose

Inputs: poses: (3, 4) bds: bounds x: translational perturb range angle: rotation angle perturb range in degrees

Inputs: poses: (3, 4) bds: bounds x: translational perturb range angle: rotation angle perturb range in degrees

Inputs: poses: (3, 4) x: translational perturb range angle: rotation angle perturb range in degrees Outputs:

processes the 1x12 raw pose from dataset by aligning and then normalizing produce logq :param poses_in: N x 12 :param mean_t: 3 :param std_

processes the position + quaternion raw pose from dataset to position + rotation matrix :param poses_in: N x 7 :param rot_mat: N x 3 x 3 :ret

Applies logarithm map to q :param q: (4,) :return: (3,)

Implementation of original NeRF seems like it is different with rendering implementation from nerf-w https://github.com/kwea123/nerf_pl/b

Remove some overlap data in val set so that train set and val set do not have overlap

Remove some overlap data in val set so that train set and val set do not have overlap

Sample a random valid homography. Computes the homography transformation between a random patch in the original image and a warped projection

select nearest_neighbor views based on unique_frame_index

image_pred and image_gt: (1, 3, H, W)

print 1 pair of salient feature map

print 1 pair of batch of salient feature map :param: target_in [B, 3, H, W] :param: rgb_in [B, 3, H, W] :param: features_target [B, C

print 1 pair of batch of salient feature map :param: target_in [B, 3, H, W] :param: rgb_in [B, 3, H, W] :param: features_target [B, C

print 1 pair of 1 sample of salient feature map :param: target_in [B, 3, H, W] :param: rgb_in [B, 3, H, W] :param: features_target [B

finetune pretrained PoseNet using NeRF

Perform 1 step of training using scheme1

naive implementation of triplet loss :param criterion: loss function :param f1: [lvl, B, C, H, W] :param f2: [lvl, B, C, H, W] :

triplet loss with hard negative mining, inspired by http://www.bmva.org/bmvc/2016/papers/paper119/paper119.pdf section3.3 :param criterion:

visualize predicted pose result vs. gt pose