Embedded Files
The detailed feature of HeLiPR Dataset
HeLiPR utilizes four different LiDARs (Two of Spinning and Two of Solid State)
The hardware configurations can introduce various challenges associated with resolution, field of view (FOV), and scanning patterns. For instance:
Ouster and Velodyne differ in resolution, with 128 channels versus 16 channels, respectively.
Ouster and Aeva have contrasting FOVs, with the former providing 360 degrees and the latter 120 degrees.
The scanning patterns of Ouster and Livox are distinct as well, with Ouster employing a repetitive pattern and Livox a non-repetitive one.
The HeliPR dataset offers a variety of channels, including x, y, z, intensity, reflectivity, NIR, and Radial Velocity.
These channels can be leveraged to differentiate objects with similar geometric features. Notably, velocity can be particularly useful in place recognition tasks.
The sequences in the HeLiPR dataset were captured in various locations, highlighting spatial variance, over a one-month period to account for temporal variance.
Given that the dataset encompasses three distinct locations, it offers environments suitable for evaluating the generalization capabilities of place recognition algorithms.
Moreover, each location is revisited three times via different routes, allowing for an assessment of the robustness of place recognition algorithms.
Each sequence in the HeLiPR dataset presents its own challenges due to inherent environmental features.
Each path possesses challenges such as various rotation variations, FOV issues due to narrow alleys, dynamic objects, and repeated occurrences of similar scenes.
The HeLiPR Dataset provides 6-DoF ground truth for each LiDAR.
The ground truth for each LiDAR undergoes a spatial transformation based on the INS using extrinsic calibration. It provides the position corresponding to the timestamp of each LiDAR measurement, determined using B-Spline Interpolation.
The HeLiPR's sequence environments and trajectories
Vistors can Download HeLiPR in here.
The trajectory that aligns with the aerial image is denoted as the 01 index in all sequences, such as Roundabout01, Town01, and Bridge01. The trajectory features a gradient color, starting with red at the beginning point and transitioning to blue at the end point.
Additionally, the trajectories for sequences 01, 02, and 03 are displayed in the third row, represented over time on the z-axis.
Characteristics of Sequences
Roundabout: This sequence features three roundabouts, each delineated by a colored box. A prominently large roundabout, paired with an external hexagon, facilitates straightforward revisiting of previously traversed locations.
Town: This sequence showcases a juxtaposition of tight alleyways and expansive boulevards. Zones labeled as A, B, and C span approximately 15m, 3m, and 5m in width, respectively. Consequently, utilizing the wide scanning potential of LiDAR can pose a challenge in the constricted alleys. The sequence is also populated with dynamic elements, including pedestrians (represented by blue points) and vehicles (indicated by red points).
Bridge: This sequence introduces a unique set of challenges, primarily stemming from the scene similarities contained within the trajectory. Sections of the bridge outlined in blue and green boxes bear a striking resemblance to one another. However, nuanced distinctions are present, as highlighted in the red box. Additionally, achieving accurate place recognition becomes intricate due to the presence of dynamic entities and the fluctuating quantities of such objects. The accompanying image vividly displays a dynamic object (marked in red points) positioned at both Bridge01 and Bridge02.
All of maps and partial maps are constructed by MA-LIO, which can find in here.
The trajectory of sequences in HeLiPR Dataset
All the trajectories have been transformed to align with the 01 sequence, where the starting location of the 01 sequence serves as the origin point (axis : meter).
The size of LiDARs are represented individually. (O : OS2-128, A : Aeva Aeries II, V : VLP-16, L : Livox Avia)
Bridge consists of 4 sequences with driving in normal (02, 03) and in reverse (01, 04)
Roundabout01
Time : Night / 2023-07-16 (2730 sec)
Location : Ansan
Complexity : ★
Path Length : 9040m
Size : O (79GB), A (54GB), V (11GB), L (8.6 GB)
Roundabout02
Time : Daytime / 2023-08-01 (2085 sec)
Location : Ansan
Complexity : ★★★
Path Length : 7447m
Size : O (58GB), A (47GB), V (8.4GB), L (6.6 GB)
Roundabout03
Time : Daytime / 2023-08-13 (2515 sec)
Location : Ansan
Complexity : ★★
Path Length : 9262m
Size : O (72GB), A (57GB), V (9.4GB), L (8.0GB)
Town01
Time : Evening / 2023-07-18 (2414 sec)
Location : Seochon
Complexity : ★★★
Path Length : 7832m
Size : O (70GB), A (55GB), V (11GB), L (8.1GB)
Town02
Time : Daytime / 2023-07-31 (2689 sec)
Location : Seochon
Complexity : ★★★★
Path Length : 8203m
Size : O (78GB), A (61GB), V (12GB), L (8.9GB)
Town03
Time : Night / 2023-08-14 (2528 sec)
Location : Seochon
Complexity : ★★★★
Path Length : 8903m
Size : O (74GB), A (63GB), V (11GB), L (8.2GB)
Bridge01
Time : Night / 2023-07-17 (2144 sec)
Location : Hangang & Dongjak Bridge
Complexity: ★★
Path Length : 23056m
Size : O (56GB), A (34GB), V (7.3GB), L (6.0GB)
Features : Fast driving (~90km/h) and driving in reverse
Bridge02
Time : Daytime / 2023-07-31 (2562 sec)
Location : Hangang & Dongjak Bridge
Complexity: ★★★★
Path Length : 14615m
Size : O (61GB), A (46GB), V (8.7GB), L (7.2GB)
Features : A lot of dynamic objects (especially vehicle) and slow driving
Bridge03
Time : Night / 2023-08-14 (2009 sec)
Location : Hangang & Dongjak Bridge
Complexity: ★★★
Path Length : 19400m
Size : O (52GB), A (36GB), V (6.9GB), L (5.6GB)
Features : Some of dynamic objects and normal driving (~60km/h)
Bridge04
Time : Daytime / 2023-08-21 (3033 sec)
Location : Hangang & Dongjak Bridge
Complexity: ★★★
Path Length : 22958m
Size : O (75GB), A (61GB), V (11GB), L (8.2GB)
Features : Some of dynamic objects (especially vehicle) and driving in reverse
The trajectory of sequences in HeLiPR Dataset with MulRan
For long-term place recognition, we also provide the three sequences with three places, KAIST, DCC and Riverside.
Below image shows trajectories derived from GPS and INS measurements. Four trajectories for each location and changes between Dajeon03-06 are represented.
We achieved sequence04 and 05 at night and morning. Furthermore, sequence06 is acheived at night after four months from sequence 04-05.
Original sequence 01-03 from MulRan are achieved from OS1-64; however, sequence04-06 from HeLiPR contains the aforementioned measurements without OS1-64.
Sequence04 does not contain IMU measurements; however it does not effect for inter-LiDAR place recognition.
KAIST04
Time : Night / 2023-08-31 (1261 sec)
Location : KAIST campus
Complexity: ★
Path Length : 6348m
Size : O (36GB), A (25GB), V (4.4GB), L (3.6GB)
Features : Few dynamic objects with long-term differences
KAIST05
Time : Morning / 2023-08-31 (1248 sec)
Location : KAIST campus
Complexity: ★★★
Path Length : 6878m
Size : O (34GB), A (28GB), V (4.7GB), L (3.7GB)
Features : Many dynamic objects such as people with long-term differences
KAIST06
Time : Night / 2024-01-16 (1215 sec)
Location : KAIST campus
Complexity: ★★★
Path Length : 6661m
Size : O (35GB), A (24GB), V (5.1GB), L (4.3GB)
Features : Few dynamic objects with long-term differences with KAIST04-05
DCC04
Time : Night / 2023-08-31 (786 sec)
Location : Dajeon Convention Center
Complexity: ★
Path Length : 5506m
Size : O (22GB), A (15GB), V (2.8GB), L (2.6GB)
Features : Few dynamic with long-term differences, inverse route with DCC03
DCC05
Time : Morning / 2023-08-31 (1081 sec)
Location : Daejeon Convention Center
Complexity: ★
Path Length : 5309m
Size : O (31GB), A (23GB), V (4.2GB), L (3.5GB)
Features : Few dynamic with long-term differences
DCC06
Time : Night / 2024-01-16 (1074 sec)
Location : Daejeon Convention Center
Complexity: ★
Path Length : 4648m
Size : O (31GB), A (21GB), V (4.4GB), L (4.1GB)
Features : Few dynamic objects with long-term differences with DCC04-05
Riverside04
Time : Night / 2023-08-31 (612 sec)
Location : KAIST and Daedeok Bridge
Complexity: ★★
Path Length : 6523m
Size : O (15GB), A (8.4GB), V (1.8GB), L (1.7GB)
Features : Few dynamic with long-term differences, High speed (~60km/h)
Riverside05
Time : Morning / 2023-08-31 (855 sec)
Location : KAIST and Daedeok Bridge
Complexity: ★★★
Path Length : 6394m
Size : O (20GB), A (15GB), V (2.9GB), L (2.2GB)
Features : Many dynamic objects with long-term differences
Riverside06
Time : Night / 2024-01-16 (1195 sec)
Location : KAIST and Daedeok Bridge
Complexity: ★★
Path Length : 7219m
Size : O (28GB), A (18GB), V (4.3GB), L (3.7GB)
Features :Few dynamic objects with long-term differences with Riverside04-05 (Low speed)
Page updated
Google Sites
Report abuse