EastPark

NOTE: to visualize the data, use the public method https://github.com/PRBonn/semantic-kitti-apimodify the code to reshape to Nx8 instead of Nx4(at auxiliary/laserscan.py:74

scan = scan.reshape((-1, 4))

should become->

scan = scan.reshape((-1, 8))

)Robotic navigation in outdoor environments requires not only semantic understanding of objects but also detailed characterization of terrain for safe navigation. Existing LiDAR datasets mostly focus on structured urban scenes, leaving unaddressed the challenges of crowded public spaces with diverse ground types. We present a new large-scale multi-sensor dataset collected in a public park using a pitched sensor setup. In addition to over 9,000 annotated LiDAR scans, the dataset also includes more than 13,000 camera images captured with a Kinect Azure and high-frequency LiDAR-integrated IMU data, providing complementary multi-modal data. Each point cloud has been manually annotated by multiple expert annotators into 22 semantic classes covering terrain, vegetation, infrastructure, and dynamic agents. Unlike previous datasets, EastPark is designed as a dedicated benchmark for semantic segmentation in complex park environments. To establish baselines, we adapt several state-of-the-art 3D semantic segmentation models using parameter-efficient fine-tuning such as linear probing and low rank adaptation (LoRA), and release all trained models alongside the dataset. This benchmark provides the community with a new tool to study semantic segmentation, as well as efficient adaptation techniques in robotics perception.