/deep_sdf_debug

Primary LanguagePython

PPT:

  1. Full point cloud - car suraface at the origin
  2. show reconstruction
  3. next step - put points at a transform
  4. goal is to optimize even now, estimate transform
  5. show visualization - mention need for good initialization

install environment

./install.sh

Sliding Window Optimization

python ours/sliding_window_reconstruct_multiple_frame_argo.py --config configs/config_kitti.json --sequence_dir data/P04/cleaned_data/000/000003/pcd.npy

Deep_SDF in DSP-SLAM default

python dsp_slam/reconstruct_frame.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 0
python dsp_slam/reconstruct_frame.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 50
python dsp_slam/reconstruct_frame.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 100
python dsp_slam/reconstruct_frame.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 500
python dsp_slam/reconstruct_frame.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 1000

Deep_SDF in DSP-SLAM without render term

python dsp_slam/reconstruct_frame_no_render_term.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 0
python dsp_slam/reconstruct_frame_no_render_term.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 50
python dsp_slam/reconstruct_frame_no_render_term.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 100
python dsp_slam/reconstruct_frame_no_render_term.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 500
python dsp_slam/reconstruct_frame_no_render_term.py --config configs/config_kitti.json --sequence_dir data/dsp_slam/kitti/07 --frame_id 1000

Jupyter notebook for cleaning data and extract to pcd extension

  1. Run all with table.
python3 ours/run_all.py
  1. Run all bash no table
./run_all.sh
  1. Evaluate optimization
python ours/2_reconstruct_multiple_frame_argo_compare.py --config configs/config_kitti.json --sequence_dir data/P04/cleaned_data/000/000009/pcd.npy

python ours/2_reconstruct_multiple_frame_argo_compare.py --config configs/config_kitti.json --sequence_dir data/P04/cleaned_data/002/002001/pcd.npy


# motion
# Pose esimation/ nosiy
#  adn reconstruction chamfer distance
# inconsidtent over time and no using for data of mulitple frame
# ablation first frame vs multiple frame
# methodology tracking optimization, kiss icp pv-rcnn deepsdf
# abstraction -> Briftly introduction 200 words
# 1. introduction: why? what? how? *Claims
# 2. Related works: mention the exising works
# 3. methodology: how we do that this
# 4. Experiemtn: explain result, desscribe dataset, evaluatuoin metric, baseline(yue pan paper), implementation deitail(parameter setting, learnng, ) show result. abltion, change some params -> should support claim
# 5. Conclusion: repeat abstraction and disadvatange , futre work how can we imrpove

Shashank:

  1. The naming of pts_cam_pcd, T_cam_obj seems to be wrong. velodyne in place of cam seems correct to me.
  2. [NOT TRUE] The surface_points and pts_cam_pcd are not with a relative transform of T_cam_obj. small difference. Visible with car_no = 1; frame = 10