Skip to content

Latest commit

 

History

History
81 lines (53 loc) · 3.13 KB

LBA.md

File metadata and controls

81 lines (53 loc) · 3.13 KB
Run DSO to get the pose esimations & Run inference with Local Bundle Adjustment

To run on the demo data (7scenes office seq 1):

(1) Run DSO to get the initial poses

python run_dso.py --dso_path ../third_party/dso/build/bin/dso_dataset \
                  --data_fldr ../data/7scenes_office_seq_01 \
                  --cam_info_file DSO/cam_info_7scenes.mat \
                  --res_path ./dso_res \
                  --name_pattern *.color* \
                  --maxframe 500

The parameters for run_dso.py:

dso_path : the built executable for DSO

data_fldr : the image folder

name_pattern : the image name pattern, e.g. *.color.png for frame_#.color.png for 7scene and scanNet name pattern

res_path : the folder path to save the dso results

maxframe : the maximal # of frames we will deal with

(2) Test KV net with LBA

python3 test_KVNet_LBA.py \
 		--exp_name LBA_demo \
 		--sigma_soft_max 10 \
 		--t_win 2 \
 		--d_min .1  --d_max 5. \
 		--feature_dim 64  --ndepth 64 \
 		--dataset single_folder --dataset_path ../data/7scenes_office_seq_01 \
        --img_name_pattern *.color.png \
        --intrin_path ./DSO/cam_info_7scenes.mat \
 		--LBA_max_iter 20  --LBA_step 0.01 \
 		--opt_r 0  --opt_t 1 \
        --use_dso_R  \
 		--opt_next_frame \
 		--min_frame_idx 0  --max_frame_idx 498 \
 		--refresh_frames 200 \
		--model_path  ./saved_models/kvnet_scannet.tar

The parameters for test_KVNet_LBA_.py

exp_name: the name for the experiment. The results will be saved in ../results/${exp_name}

sigma_soft_max: the sigma value used for getting the DPV. Should be the same as during training.

t_win: the time window radius. The time window will be centered around the current frame

d_min, d_max : the minimal and maximal depth values

feature_dim: PSM feature dimension in DNet, should be the same as in the training session

dataset: the dataset name. Should one of scanNet, 7scenes, kitti, single_folder. For single folder, the dataset_path is the path of the folder that includes the input rgb images

dataset_path: the path to the specified dataset

model_path: the path to the trained model

img_name_pattern : used for the case where dataset == single_folder. The image name pattern, e.g. *.color.png for framecolor.png for 7scene and scanNet name pattern

intrin_path: the .mat file including the camera intrinsic

LBA_max_iter, LBA_step : the maximal iteration and step size for the optimization in LBA

opt_r, opt_t: whether optimizing rotation R and translation. 0: not optimize (use initial guess, if there is any). 1: optimize

use_dso_R, use_dso_t: whether to use R or t from DSO as the initializations

opt_next_frame: also optmize the the frame right next to the current frame. This is benificial for propagating the depth probability volume and getting stable results

refresh_frame: refreshing' the pose estimation every ${refresh_frame} frames. By 'refreshing', we mean resetting the pose estimation to the DSO estimation. This is a simple hack for dealing with the drifting issue.

min_frame_idx, max_frame_idx : the min/max index for the images we are going to test