Add MAN TruckScenes data processing into codebase

README.md

@@ -102,7 +102,7 @@ If you prefer to build the Docker image by yourself, Check [build-docker-image](
 
 ## 1. Data Preparation
 
-Refer to [dataprocess/README.md](dataprocess/README.md) for dataset download instructions. Currently, we support **Argoverse 2**, **Waymo**, **nuScenes**, **ZOD** and **custom datasets** (more datasets will be added in the future). 
+Refer to [dataprocess/README.md](dataprocess/README.md) for dataset download instructions. Currently, we support **Argoverse 2**, **Waymo**, **nuScenes**, [**MAN-TruckScene**](https://github.com/TUMFTM/truckscenes-devkit), [**ZOD**](https://github.com/zenseact/zod) and **custom datasets** (more datasets will be added in the future). 
 
 After downloading, convert the raw data to `.h5` format for easy training, evaluation, and visualization. Follow the steps in [dataprocess/README.md#process](dataprocess/README.md#process). 
 
@@ -337,6 +337,12 @@ If you find it useful, please cite our works:
     year={2025},
     journal={arXiv preprint arXiv:2503.00803},
 }
+@article{zhang2025deltaflow,
+    title={{DeltaFlow}: An Efficient Multi-frame Scene Flow Estimation Method},
+    author={Zhang, Qingwen and Zhu, Xiaomeng and Zhang, Yushan and Cai, Yixi and Andersson, Olov and Jensfelt, Patric},
+    year={2025},
+    journal={arXiv preprint arXiv:2508.17054},
+}
 ```
 
 And our excellent collaborators works contributed to this codebase also:

assets/slurm/data-process.sh

@@ -0,0 +1,44 @@
+#!/bin/bash
+#SBATCH -J dataprocess
+#SBATCH -p berzelius-cpu
+#SBATCH --cpus-per-task 64
+#SBATCH --mem 128G
+#SBATCH --mincpus=64
+#SBATCH -t 2-00:00:00
+#SBATCH --mail-type=END,FAIL
+#SBATCH --mail-user=qingwen@kth.se
+#SBATCH --output /proj/berzelius-2023-154/users/x_qinzh/OpenSceneFlow/logs/slurm/%J_data.out
+#SBATCH --error  /proj/berzelius-2023-154/users/x_qinzh/OpenSceneFlow/logs/slurm/%J_data.err
+
+PYTHON=/proj/berzelius-2023-154/users/x_qinzh/mambaforge/envs/sftool/bin/python
+export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/proj/berzelius-2023-154/users/x_qinzh/mambaforge/lib
+cd /proj/berzelius-2023-154/users/x_qinzh/OpenSceneFlow/
+# export HYDRA_FULL_ERROR=1
+
+
+# =============== Argoverse2
+# $PYTHON dataprocess/extract_av2.py --nproc 64 \
+#     --av2_type sensor \
+#     --data_mode train \
+#     --argo_dir /proj/berzelius-2023-154/users/x_qinzh/av2 \
+#     --output_dir /proj/berzelius-2023-364/users/x_qinzh/data/av2/h5py
+
+# $PYTHON dataprocess/extract_av2.py --nproc 64 \
+#     --av2_type sensor \
+#     --data_mode val \
+#     --argo_dir /proj/berzelius-2023-154/users/x_qinzh/av2 \
+#     --output_dir /proj/berzelius-2023-364/users/x_qinzh/data/av2/h5py \
+#     --mask_dir /proj/berzelius-2023-154/users/x_qinzh/av2/3d_scene_flow
+
+# $PYTHON dataprocess/extract_av2.py --nproc 64 \
+#     --av2_type sensor \
+#     --data_mode test \
+#     --argo_dir /proj/berzelius-2023-154/users/x_qinzh/av2 \
+#     --output_dir /proj/berzelius-2023-364/users/x_qinzh/data/av2/h5py \
+#     --mask_dir /proj/berzelius-2023-154/users/x_qinzh/av2/3d_scene_flow
+
+# =============== MAN-Truckscenes
+$PYTHON dataprocess/extract_truckscenes.py --nproc 12 \
+    --data_dir /proj/berzelius-2023-364/users/x_qinzh/data/truckscenes/raw_data/trainval/man-truckscenes \
+    --mode v1.0-trainval \
+    --output_dir /proj/berzelius-2023-364/data/truckscenes/h5py

assets/tests/hdbscan_speed.py

@@ -25,7 +25,7 @@
 
 def vis(pc0, labels, title='HDBSCAN'):
     # visualize
-    vis = MyVisualizer(view_file=f'{BASEF_DIR}/assets/view/av2.json', window_title=title)
+    vis = MyVisualizer(view_file=f'{BASEF_DIR}/assets/view/demo.json', window_title=title)
     pcd = o3d.geometry.PointCloud()
     num_points = pc0.shape[0]
     pcd.points = o3d.utility.Vector3dVector(pc0)

assets/view/demo.json

@@ -0,0 +1,19 @@
+{
+	"class_name" : "ViewTrajectory",
+	"interval" : 29,
+	"is_loop" : false,
+	"trajectory" : 
+	[
+		{
+			"boundingbox_max" : [ 215.05035400390625, 80.301795959472656, 34.890830993652344 ],
+			"boundingbox_min" : [ -211.76950073242188, -125.19444274902344, -4.2829980850219727 ],
+			"field_of_view" : 90.0,
+			"front" : [ -0.58999680455306347, -0.15524870408634167, 0.79233932787454509 ],
+			"lookat" : [ 9.9369846496341676, 13.322769196068615, -16.164870152242376 ],
+			"up" : [ 0.79499677526385637, 0.059685607254847638, 0.60367023746967163 ],
+			"zoom" : 0.17999999999999999
+		}
+	],
+	"version_major" : 1,
+	"version_minor" : 0
+}

conf/others/truckscenes.toml

@@ -0,0 +1,31 @@
+
+[important]
+height = 0        # ego height. since truckscene have multiple LiDARs, we set pc to ego-coord already.
+
+# 整个雷点点云以自己为中心 分为多少个segment，每个segment又分成多少个bin
+[segments]
+r_min = 0.1          # minimum point distance. 感兴趣的区域
+r_max = 100           # maximum point distance.
+n_segments = 360     # number of radial segments.
+n_bins = 160         # number of radial bins.
+
+
+[ground]
+# 直线的 slope 斜率，一个seg里拟合的线斜率
+min_slope = 0.0      # minimum slope of a ground line. [T_m_small]
+max_slope = 0.1     # maximum slope of a ground line. [T_m]
+# 如果拟合点超过这个 long threshold，判断其 height 变化是否超过 max_long_height
+long_threshold = 2.0    # Distance at which points are considered far from each other. [T_d_prev]
+max_long_height = 0.3   # maximum height change to previous point in long line. [T_b]
+# 如果没有超过 long threshold，判断现在拟合的点与fitline高度变化是否超过 max_start_height
+max_start_height = 0.2  # Maximum heigh of starting line to be labelled ground.
+# 最大的拟合误差，超过这个误差的点不会被加入 line 拟合
+max_fit_error = 0.1  # maximum error of a point during line fit. [T_RMSE: will sqaure inside code.]
+
+# 判断是否是地面点的时候，需要满足的条件：
+max_dist_to_line = 0.2   # maximum vertical distance of point to line to be considered ground. [T_d_ground]
+line_search_angle = 2.0 # How far to search for a line in angular direction [rad].
+
+[general]
+n_threads = 8        # number of threads for parallel processing.
+verbose = false       # if you don't want to see every output, set this to false.

conf/ground/zod.toml → conf/others/zod.toml

@@ -23,7 +23,7 @@ max_start_height = 0.2  # Maximum heigh of starting line to be labelled ground.
 max_fit_error = 0.1  # maximum error of a point during line fit. [T_RMSE: will sqaure inside code.]
 
 # 判断是否是地面点的时候，需要满足的条件：
-max_dist_to_line = 0.2   # maximum vertical distance of point to line to be considered ground. [T_d_ground]
+max_dist_to_line = 0.25   # maximum vertical distance of point to line to be considered ground. [T_d_ground]
 line_search_angle = 2.0 # How far to search for a line in angular direction [rad].
 
 [general]

dataprocess/README.md

@@ -6,13 +6,13 @@ README for downloading and preprocessing the dataset. We includes waymo, argover
 - [Download](#download): includes how to download the dataset.
 - [Process](#process): run script to preprocess the dataset.
 
-We've updated the process dataset for:
+We've updated the process dataset for (Please cite the original dataset paper and involved work if you use them):
 
 - [x] Argoverse 2.0: check [here](#argoverse-20). The process script Involved from [DeFlow](https://github.com/KTH-RPL/DeFlow).
 - [x] Waymo: check [here](#waymo-dataset). The process script was involved from [SeFlow](https://github.com/KTH-RPL/SeFlow).
 - [x] nuScenes: check [here](#nuscenes), The process script was involved from [DeltaFlow](https://github.com/Kin-Zhang/DeltaFlow).
 - [x] ZOD (w/o gt): check [here](#zod-dataset). The process script was involved from [HiMo](https://kin-zhang.github.io/HiMo). (It could be a good first reference for users to extract other datasets in the future.)
-- [ ] TruckScene: done coding, public after review. Will be involved later by another paper.
+- [x] TruckScene: check [here](#truckscene). The process script was involved from [DoGFlow](https://github.com/ajinkyakhoche/DoGFlow).
 
 If you want to **use all datasets above**, there is a **specific environment** in [envsftool.yaml](../envsftool.yaml) to install all the necessary packages. As Waymo package have different configuration and conflict with the main environment. Setup through the following command:
 
@@ -134,6 +134,46 @@ For HiMo, we only downloaded [drives-set](https://zod.zenseact.com/drives/) for
 
 Please check the scripts: [dataprocess/extract_zod.py](./extract_zod.py) in detail, current we only process one scene while feel free to comment out for all scene etc.
 
+### TruckScene
+
+Please visit the [TruckScene dataset](https://brandportal.man/d/QSf8mPdU5Hgj/downloads#/-/dataset) page for privacy policy. You can download the dataset by following command:
+
+```bash
+# mini set, ~11G recommended for debugging purpose
+cd /home/kin/data/truckscene/mini
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/mini/man-truckscenes_metadata_v1.0-mini.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/mini/man-truckscenes_sensordata_v1.0-mini.zip
+unzip "man-truckscenes_*.zip"
+
+# full trainval set, ~630G
+cd /home/kin/data/truckscene/trainval
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_metadata_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata01_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata02_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata03_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata04_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata05_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata06_v1.0-trainval.zip
+wget https://man-truckscenes.s3.eu-central-1.amazonaws.com/release/trainval/man-truckscenes_sensordata07_v1.0-trainval.zip
+unzip "man-truckscenes_*.zip"
+```
+
+Folder structure:
+```
+truckscene/
+  — samples/
+  — sweeps/
+  — v1.0-mini/
+  — v1.0-trainval/
+```
+
+#### Dataset frames
+
+| Dataset | # Total Scene | # Total Frames |
+| ------- | ------------- | -------------- |
+| train   | 524           | 101902 / 20380 (w. gt)         |
+| val     | 75           | 14625 / 2925 (w. gt)          |
+
 
 
 ## Process
@@ -144,8 +184,9 @@ This directory contains the scripts to preprocess the datasets into `.h5` files.
 - `extract_nus.py`: Process the datasets in nuScenes.
 - `extract_waymo.py`: Process the datasets in Waymo.
 - `extract_zod.py`: Process the datasets in ZOD.
+- `extract_truckscene.py`: Process the datasets in TruckScene.
 
-Example Running command:
+Example Running command, you can also check our [slurm data-process script](../assets/slurm/data-process.sh) for more details.:
 ```bash
 # av2:
 python dataprocess/extract_av2.py --av2_type sensor --data_mode train --argo_dir /home/kin/data/av2 --output_dir /home/kin/data/av2/h5py
@@ -155,6 +196,9 @@ python dataprocess/extract_waymo.py --mode train --flow_data_dir /home/kin/data/
 
 # nus:
 python dataprocess/extract_nus.py --mode v1.0-trainval --output_dir /home/kin/data/nus/h5py/full --nproc 24
+
+# truckscene:
+python dataprocess/extract_truckscene.py --data_dir /home/kin/data/man-truckscenes --mode v1.0-mini --output_dir /home/kin/data/nus/h5py
 ```
 
 

dataprocess/extract_nus.py

@@ -18,7 +18,7 @@
 import multiprocessing
 from pathlib import Path
 from multiprocessing import Pool, current_process
-from typing import Optional, Final
+from typing import Optional
 from tqdm import tqdm
 import numpy as np
 import fire, time, h5py
@@ -69,10 +69,13 @@ def _load_points_from_file(filename: str) -> np.ndarray:
     pc = pc.reshape((-1, 5))[:, :4]
     return pc
 
-def get_pose(nusc, sweep_data):
-    # pose from lidar to world
-    ego2lidar = nusc.get('calibrated_sensor', sweep_data['calibrated_sensor_token'])
-    world2ego = nusc.get('ego_pose', sweep_data['ego_pose_token'])
+def get_pose(data_fn, sweep_data, w2stf=True):
+    world2ego = data_fn.get('ego_pose', sweep_data['ego_pose_token'])
+    # without considering the sensor to ego transform, we will do it outside if multiple sensors are used.
+    if not w2stf:
+        return transform_matrix(world2ego['translation'], Quaternion(world2ego['rotation'])).astype(np.float64)
+
+    ego2lidar = data_fn.get('calibrated_sensor', sweep_data['calibrated_sensor_token'])
     ego2lidar_np = transform_matrix(ego2lidar['translation'], Quaternion(ego2lidar['rotation']))
     world2ego_np = transform_matrix(world2ego['translation'], Quaternion(world2ego['rotation']))
     return np.dot(world2ego_np, ego2lidar_np)
@@ -85,13 +88,13 @@ def if_annotated_frame(sample_ann_dict, ts0):
             break
     return gt_flow_flag
 
-def _resample_data(nusc, sample_data, sample_ann_dict, resample2frequency=10):
+def _resample_data(nusc, sample_data, sample_ann_dict, datafrequency=20, resample2frequency=10):
     """
     NOTE(Qingwen) - 2025-05-18:
     We always want to start from the first GT frame, and then resample the data! So we have as many GT frames as possible...
     """
     sweep_data_lst, timestamps_lst = [], []
-    skipFrame = int(20 / resample2frequency)  # since nuscenes sweep at 20Hz, we want to resample to 10Hz
+    skipFrame = int(datafrequency / resample2frequency)  # since nuscenes sweep at 20Hz, we want to resample to 10Hz
     cnt = 0
 
     # Find the first GT frame
@@ -268,18 +271,17 @@ def proc(x, ignore_current_process=False):
         pos = 1
     process_log(*x, n=pos)
 
-def process_logs(nusc_mode, data_dir: Path, scene_list: list, output_dir: Path, nproc: int):
-    """Compute sceneflow for all logs in the dataset. Logs are processed in parallel.
-       Args:
-         data_dir: Argoverse 2.0 directory
-         output_dir: Output directory.
+def process_logs(data_mode, data_dir: Path, scene_list: list, output_dir: Path, nproc: int):
     """
-
+    Compute sceneflow for all logs in the dataset. 
+    Logs are processed in parallel.
+    """
+
     if not (data_dir).exists():
         print(f'{data_dir} not found')
         return
 
-    args = sorted([(nusc_mode, data_dir, scene_num_id, output_dir) for scene_num_id in range(len(scene_list))])
+    args = sorted([(data_mode, data_dir, scene_num_id, output_dir) for scene_num_id in range(len(scene_list))])
     print(f'Using {nproc} processes')
 
     # # for debug
@@ -295,7 +297,7 @@ def process_logs(nusc_mode, data_dir: Path, scene_list: list, output_dir: Path,
             res = list(tqdm(p.imap_unordered(proc, args), total=len(scene_list), ncols=100))
 
 def main(
-    data_dir: str = "/home/kin/data/nus/v1.0",
+    data_dir: str = "/home/kin/data/nus/raw",
     mode: str = "v1.0-mini",
     output_dir: str ="/home/kin/data/nus/h5py/demo",
     nproc: int = (multiprocessing.cpu_count() - 1),