AI4701-ComputerVision-3D-Recon Final Project

setup

pip install -r requirements.txt

overview

images/: 11 张图像，0000 作为世界坐标系。

main.py: 主函数

config/config.yaml: 配置文件

GLOB:
  intrinsic_path: camera_intrinsic.txt
  world_cam_path: images/0000.png
  save_dir: output

RECON_INIT:
  camera_paths: [images/0000.png, images/0001.png]

RECON:
  method: epipolar # pnp or epipolar
  visualize: True
  merge_3d: False
  normalize_epi: True # whether scale the transition vector derived from epipolar, 否则会有尺度问题

ESTIMATOR:
  alg: ransac # ransac, magsac, use in findEssentialMat, findFundamentalMat, not in match
  ransac_params:
    ransacReprojThreshold: 0.1
    confidence: 0.99
  extract:
    contrast_thresh: 0.001 # Threshold for keypoint selection based on contrast. Lower values increase feature count but reduce stability.
    edge_thresh: 10 # Threshold for eliminating edge responses in keypoints, lower value tends to ignore more features near edges, thereby reducing mismatches caused by edges
    sigma: 1.6
  match:
    thres: 0.5
    alg: None # ransac, magsac, None
    ransac_params:
      ransacReprojThreshold: 10
      confidence: 0.99
    tree: 7 # 配置索引，密度树的数量为5
    checks: 50 # 指定递归次数
    flan_k: 2 # 最近邻的数量，=2,表示寻找两个最近邻，一般不动

BA:
  least_square_params:
    method: trf # trf or lm
    ftol: 1e-8

src/

feature_extraction.py : SIFT 图像特征提取
feature_matching.py : FLAN 图像特征匹配
initial_recon.py : 对极几何三维场景初始化
pnp_recon.py : PnP 方法三维重建
bundle_adjustment.py: BA 优化
epipolar_recon.py : 仅用对极几何进行三维重建
utils.py: 工具

outputs/ : 输出路径

view/view.json ：ctrl + C 复制之后，当 o3d 窗口出现时，ctrl + V 移动到固定视角

报告中各实验结果下载 (5.35GB) ：https://pan.baidu.com/s/1U2KEm_XWg_psUyz6p5gW3Q?pwd=a943 提取码: a943

解压到 outputs/ 下

huskydoge/AI4701-ComputerVision-3D-Recon

AI4701-ComputerVision-3D-Recon Final Project

setup

overview