[MINOR] update readme on PIN-SLAM with colorized lidar points

README.md

@@ -132,7 +132,7 @@ conda activate pin
 conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.7 -c pytorch -c nvidia 
 ```
 
-The commands depend on your CUDA version. You may check the instructions [here](https://pytorch.org/get-started/previous-versions/).
+The commands depend on your CUDA version (check it by `nvcc --version`). You may check the instructions [here](https://pytorch.org/get-started/previous-versions/).
 
 ### 3. Install other dependency
 
@@ -175,6 +175,11 @@ Use `run_demo_sem.yaml` if you want to conduct metric-semantic SLAM using semant
 python3 pin_slam.py ./config/lidar_slam/run_demo_sem.yaml -vsm
 ```
 
+Use `run_kitti_color.yaml` if you want to test PIN-SLAM with the colorized point cloud using also the RGB images:
+```
+python3 pin_slam.py ./config/lidar_slam/run_kitti_color.yaml kitti 00 -i ./data/kitti_example -vsmd
+```
+
 If you are running on a server without an X service (you may first try `export DISPLAY=:0`), then you can turn off the visualization `-v` flag:
 ```
 python3 pin_slam.py ./config/lidar_slam/run_demo.yaml -sm
@@ -256,7 +261,7 @@ The SLAM results and logs will be output in the `output_root` folder set in the
 
 For evaluation, you may check [here](https://github.com/PRBonn/PIN_SLAM/blob/main/eval/README.md) for the results that can be obtained with this repository on a couple of popular datasets. 
 
-The training logs can be monitored via Weights & Bias online if you set the flag `-w`. If it's your first time using [Weights & Bias](https://wandb.ai/site), you will be requested to register and log in to your wandb account. You can also set the flag `-l` to turn on the log printing in the terminal and set the flag `-r` to turn on the visualization logging by [rerun](https://github.com/rerun-io/rerun).
+The training logs can be monitored via Weights & Bias online if you set the flag `-w`. If it's your first time using [Weights & Bias](https://wandb.ai/site), you will be requested to register and log in to your wandb account. You can also set the flag `-l` to turn on the log printing in the terminal and set the flag `-r` to turn on the visualization logging by [rerun](https://github.com/rerun-io/rerun). If you want to get the dense merged point cloud map using the estimated poses of PIN-SLAM, you can set the flag `-p`.
 
 </details>
 

config/lidar_slam/run_kitti_color.yaml

@@ -15,10 +15,10 @@ process:
   min_z_m: -3.5
 sampler:
   surface_sample_range_m: 0.25 
-  surface_sample_n: 4
+  surface_sample_n: 3
   free_sample_begin_ratio: 0.3
   free_sample_end_dist_m: 1.0
-  free_front_sample_n: 2
+  free_front_sample_n: 4 # 2
 neuralpoints:
   voxel_size_m: 0.4
   feature_dim: 8 
@@ -35,7 +35,7 @@ loss:
   ekional_loss_on: True
   weight_e: 0.5
 continual:
-  batch_size_new_sample: 1000
+  batch_size_new_sample: 4000
   pool_capacity: 2e7
 tracker:
   source_vox_down_m: 0.6

dataset/dataloaders/apollo.py

@@ -2,7 +2,6 @@
 #
 # Copyright (c) 2022 Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Cyrill
 # Stachniss.
-# 2024 Yue Pan
 #
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -29,19 +28,12 @@
 
 import natsort
 import numpy as np
+import open3d as o3d
 from pyquaternion import Quaternion
 
 
 class ApolloDataset:
     def __init__(self, data_dir: Path, *_, **__):
-        try:
-            self.o3d = importlib.import_module("open3d")
-        except ModuleNotFoundError:
-            print(
-                'pcd files requires open3d and is not installed on your system run "pip install open3d"'
-            )
-            sys.exit(1)
-
         self.scan_files = natsort.natsorted(glob.glob(f"{data_dir}/pcds/*.pcd"))
         self.gt_poses = self.read_poses(f"{data_dir}/poses/gt_poses.txt")
         self.sequence_id = os.path.basename(data_dir)
@@ -57,7 +49,7 @@ def __getitem__(self, idx):
         return frame_data
 
     def get_scan(self, scan_file: str):
-        points = np.asarray(self.o3d.io.read_point_cloud(scan_file).points, dtype=np.float64)
+        points = np.asarray(o3d.io.read_point_cloud(scan_file).points, dtype=np.float64)
         return points.astype(np.float64)
 
     @staticmethod
@@ -82,4 +74,4 @@ def get_timestamps(points):
         y = points[:, 1]
         yaw = -np.arctan2(y, x)
         timestamps = 0.5 * (yaw / np.pi + 1.0)
-        return timestamps
+        return timestamps

dataset/dataloaders/helipr.py

@@ -34,15 +34,7 @@
 
 class HeLiPRDataset:
     def __init__(self, data_dir: Path, sequence: str, *_, **__):
-        try:
-            self.o3d = importlib.import_module("open3d")
-        except ModuleNotFoundError as e:
-            raise ModuleNotFoundError(
-                "Open3D is not installed on your system, to fix this either "
-                'run "pip install open3d" '
-                "or check https://www.open3d.org/docs/release/getting_started.html"
-            ) from e
-
+
         self.sequence_id = sequence
         self.sequence_dir = os.path.join(data_dir, "LiDAR", self.sequence_id)
         scan_files = os.listdir(self.sequence_dir)

dataset/dataloaders/kitti.py

@@ -85,6 +85,8 @@ def __init__(self, data_dir, sequence: str, *_, **__):
         # Load GT Poses (if available)
         if int(sequence) < 11:
             self.poses_fn = os.path.join(data_dir, f"poses/{self.sequence_id}.txt")
+            if not os.path.exists(self.poses_fn):
+                self.poses_fn = os.path.join(self.kitti_sequence_dir, f"poses.txt")
             self.gt_poses = self.load_poses(self.poses_fn)
 
     def __getitem__(self, idx):
@@ -226,8 +228,8 @@ def persepective_cam2image(self, points, K_mat):
         points_proj = np.matmul(K_mat[:3,:3].reshape([1,3,3]), points)
         depth = points_proj[:,2,:]
         depth[depth==0] = -1e-6
-        u = np.round(points_proj[:,0,:]/np.abs(depth)).astype(np.int32)
-        v = np.round(points_proj[:,1,:]/np.abs(depth)).astype(np.int32)
+        u = np.round(points_proj[:,0,:]/np.abs(depth)).astype(int)
+        v = np.round(points_proj[:,1,:]/np.abs(depth)).astype(int)
 
         if ndim==2:
             u = u[0]; v=v[0]; depth=depth[0]

dataset/dataloaders/kitti360.py

@@ -191,8 +191,8 @@ def persepective_cam2image(self, points, K_mat):
         points_proj = np.matmul(K_mat[:3,:3].reshape([1,3,3]), points)
         depth = points_proj[:,2,:]
         depth[depth==0] = -1e-6
-        u = np.round(points_proj[:,0,:]/np.abs(depth)).astype(np.int32)
-        v = np.round(points_proj[:,1,:]/np.abs(depth)).astype(np.int32)
+        u = np.round(points_proj[:,0,:]/np.abs(depth)).astype(int)
+        v = np.round(points_proj[:,1,:]/np.abs(depth)).astype(int)
 
         if ndim==2:
             u = u[0]; v=v[0]; depth=depth[0]

model/neural_points.py

@@ -50,10 +50,8 @@ def __init__(self, config: Config) -> None:
 
         self.device = config.device
         self.dtype = config.dtype
-        self.idx_dtype = (
-            torch.int64
-        )  # torch.int64/32 does not have much speed difference
-
+        self.idx_dtype = torch.int64
+
         self.resolution = config.voxel_size_m
 
         self.buffer_size = config.buffer_size

pin_slam.py

@@ -329,7 +329,7 @@ def run_pin_slam(config_path=None, dataset_name=None, sequence_name=None, seed=N
             if config.pgo_on:
                 print("time for loop detection and PGO (ms):", (T4-T3)*1e3)
             print("time for mapping preparation    (ms):", (T5-T4)*1e3)
-            print("time for training               (ms):", (T6-T5)*1e3)
+            print("time for mapping                (ms):", (T6-T5)*1e3)
 
         # regular saving logs
         if config.log_freq_frame > 0 and (frame_id+1) % config.log_freq_frame == 0:
@@ -392,7 +392,7 @@ def run_pin_slam(config_path=None, dataset_name=None, sequence_name=None, seed=N
                     else:
                         cur_sdf_slice = cur_sdf_slice_h
 
-            pool_pcd = mapper.get_data_pool_o3d(down_rate=17, only_cur_data=o3d_vis.vis_only_cur_samples) if o3d_vis.render_data_pool else None # down rate should be a prime number
+            pool_pcd = mapper.get_data_pool_o3d(down_rate=23, only_cur_data=o3d_vis.vis_only_cur_samples) if o3d_vis.render_data_pool else None # down rate should be a prime number
             odom_poses, gt_poses, pgo_poses = dataset.get_poses_np_for_vis()
             loop_edges = pgm.loop_edges_vis if config.pgo_on else None
             o3d_vis.update_traj(dataset.cur_pose_ref, odom_poses, gt_poses, pgo_poses, loop_edges)

requirements.txt

@@ -12,4 +12,5 @@ rich==12.5.1
 roma==1.5.0
 rospkg==1.5.1
 scikit-image==0.21.0
-wandb==0.17.0
+wandb==0.17.0
+opencv-python

scripts/download_kitti_example.sh

@@ -4,7 +4,7 @@ mkdir -p data
 cd data
 
 echo Downloading KITTI odometry dataset, sequence 00 subset, first 100 frames ...
-wget -O kitti_example.tar.gz -c https://uni-bonn.sciebo.de/s/KwOuBiPZi8vSz2O/download
+wget -O kitti_example.tar.gz -c https://uni-bonn.sciebo.de/s/Ycl28f1Cppghvjm/download
 
 echo Extracting dataset...
 tar -xvf kitti_example.tar.gz

utils/tools.py

@@ -12,6 +12,7 @@
 import subprocess
 import sys
 import time
+import warnings
 from datetime import datetime
 from pathlib import Path
 from typing import List
@@ -40,6 +41,8 @@ def setup_experiment(config: Config, argv=None, debug_mode: bool = False):
     os.environ["CUDA_VISIBLE_DEVICES"] = str(config.gpu_id)
     ts = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")  # begining timestamp
 
+    warnings.filterwarnings("ignore", category=FutureWarning) 
+
     run_name = config.name + "_" + ts  # modified to a name that is easier to index
 
     run_path = os.path.join(config.output_root, run_name)