Benchmark Results for Place Recognition

Identical LiDAR Place Recognition VS Heterogeneous LIDAR Place Recognition

0. Baseline Method with Evaluation Technique

• We utilize four methods: three for the model-based approach (Scan Context, STD and RING++  )and one for the learning-based approach (LoGG3D-Net).

• Scans from Livox Avia are typically grouped in sets of 20 due to their sparse nature. However, for the STD method, 20 scans from all LiDAR types are used. 

• All scans are downsampled to a 0.2m resolution and cropped to a 100m. This ensures a fair comparison with Velodyne, whose maximum distance is 100m.

• For evaluation, query scans are sampled at 10m intervals, while target scans are at 5m intervals. A place recognition is considered successful if a candidate is identified within a 7.5m range, which is labeled as a true positive.

• To train the LoGG3D-Net, we utilize 6 sequences with 4 LiDARs. Their scans are sampled at 1m interval. We adhered to the training schemes and parameters from the original paper, with the exceptions of adjusting the positive sampling distance to 7.5m and setting the number of negatives to 8.

• Training for LoGG3D-Net: 

We view the data acquired at the same time and space(ex. DCC05/Ouster, DCC05/Aeva, DCC05/Livox, and DCC05/Velodyne) as a single super sequence (ex. DCC05), and 6 super sequences (DCC04-05,KAIST04-05, and Riverside04-05) are selected for training data. 

Since our training datasets incorporate the diversity of scanning patterns, FOV, and dynamic object occlusion, no additional data augmentation is included. With a base learning rate of 1e-3, a multi-step learning rate scheduler (reducing by a factor of 10 every 10 epochs) and Adam optimizer with a momentum of 0.8 is utilized. For the training quadruplet, 2 positives, 9 negatives, and 1 other negative are sampled. The threshold for true positive mining is 7.5 m for both training and testing, and negatives are at least 20 m far from the positives. Due to the large size of training data, 5k query scans are randomly sampled for each epoch. Starting from scratch, training is done with a maximum of 50 epochs and early stopped when converged.

2. PR Curve for LIDAR pairs (Bridge02-03)