CropCount3D: A Multi-Modal Dataset for Real-Time 3D Crop Counting in Densely Planted Orchards

Description CropCount3D is a publicly available dataset containing synchronized camera, LiDAR, and IMU measurements acquired in densely planted orchard environments—including both greenhouses and open fields—for the purpose of on-the-fly crop counting. Designed to facilitate research in real-time 3D crop counting under challenging agricultural conditions, the dataset captures realistic field scenarios characterized by dense foliage, occlusion, and irregular planting. It includes frame-wise crop count ground truth, enabling frame-level evaluation of crop counting algorithms without requiring instance-level annotations. CropCount3D serves as a benchmark for assessing the accuracy, consistency, and robustness of multi-modal, real-time crop counting systems in unstructured environments. Sensor Platform The dataset was collected using a handheld multi-sensor rig designed for real-time crop counting in densely planted settings. The platform rigidly mounts a monocular RGB camera, a 3D LiDAR, and an IMU. All data streams are temporally synchronized and spatially calibrated. Camera: FLIR Blackfly (FL3-U3-1324C-C) with Edmund Optics 86900 lens LiDAR: Livox Avia IMU: BMI088 (built-in with LiDAR) Motion Patterns The sensor rig was manually operated through the orchard environments, following three representative motion patterns designed to reflect common real-world scanning behaviors: Revisit: Back-and-forth traversal simulating revisit scenarios prone to double-counting (3 sequences collected both in open-field and greenhouse settings) Circular: Rotational scanning around a single tree in spatially constrained environments. (2 sequences collected in greenhouse settings.) Straight: Linear forward motion, typically observed in row-wise crop arrangements. (3 sequences collected both in open-field and greenhouse settings.) The trajectory for each sequence is provided in the corresponding image files (.png). Dataset Structure The CropCount3D dataset is organized by sequence, where each sequence represents a distinct scanning session. Each sequence contains synchronized multi-sensor data and corresponding keyframe-wise cumulative crop count ground truth. Each ROS1 bag file ({$sequence_name}.bag) in the dataset/ directory contains raw sensor measurements for a specific sequence, recorded under the following topics: /camera/image_color: RGB images, 1280×1024×3 /livox/lidar: LiDAR point clouds /livox/imu: IMU data Calibration files are provided in the calibration/ directory and include the following: camera_intrinsics.yaml: Intrinsic parameters (focal length, principal point, distortion coefficients), obtained using the ROS camera_calibration tool. lidar_camera_extrinsics.yaml: 6DoF transformation from LiDAR to camera coordinate frame, obtained using the livox_camera_calib. Crop count annotations are provided in the crop_count/ directory as follows: {$sequence_name}_crop_counts.csv: Per-keyframe cumulative ground truth crop counts for each sequence. Detection weights are provided in the weight/ directory: yolov11s_seg.pt: Pretrained YOLOv11s-segmentation model weights. Compatible with Pytorch 2.0.1. yolov11s_seg.onnx: Pretrained YOLOv11s segmentation model weights, converted to ONNX format for deployment. Compatible with ONNX Runtime GPU version 1.15.0 and CUDA 11.4. Dataset Acquisition Time Data acquisition times are as follows: Revisit1: 2024-12-19 AM Revisit2: 2024-12-19 AM Revisit3: 2024-12-20 PM Circular1: 2024-01-31 PM Circular2: 2024-02-01 AM Straight1: 2024-01-31 PM Straight2: 2024-02-01 AM Straight3: 2024-12-20 AM Citation Please cite the following paper when using this dataset in your work.[ANONYMIZED FOR REVIEW] License Information The CropCount3D dataset is released under a Creative Commons Attribution 4.0 International License, CC BY 4.0 Contact Information If you have any issues about the CropCount3D dataset, please contact us at [ANONYMIZED EMAIL].