AgriChrono: A Longitudinal Dataset and Benchmark for Novel-View Synthesis of Field-Grown Crops
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "AgriChrono: A Longitudinal Dataset and Benchmark for Novel-View Synthesis of Field-Grown Crops".
Paper summary: Rosa: So to recap, this paper presents AgriChrono as a modular robotic data collection platform and a multi-modal dataset specifically designed to capture dynamic farmland conditions for three dee reconstruction of crop structures <ref:2508.18694#pg1,a modular robotic data collection platform and a multi-modal dataset>.
Dev: The thesis is that while controlled indoor environments are good for high quality object level reconstruction, scaling those methods to complex in the wild agricultural settings is really hard due to the lack of dedicated datasets and benchmarks.
Taro: AgriChrono addresses this gap by introducing a comprehensive in-the-wild multimodal dataset and benchmark suite enabled by a dedicated robotic acquisition framework designed to capture environmental variations of real agricultural scenes.
Rosa: They developed this platform for remote, hands off operation via a customized web user interface, which served as the foundation for long term time synchronized collection of high quality multimodal data including RGB, Depth, LiDAR, IMU, and Pose.
Dev: This infrastructure allows for the precise time synchronized acquisition of four view RGB two view Depth LiDAR IMU and VIO Pose all managed via a custom web user interface to maintain rigorous data protocols even from the field.
Taro: The main contribution is releasing the eighteen terabyte AgriChrono dataset which captures comprehensively in the wild environments, documenting continuous morphological changes in a complex natural setting.
Rosa: They captured this data four times per day under different illuminations during the crop’s main growth period, which provides documentation of how things change over time in a natural setting.
Dev: Furthermore, they introduce a new benchmark for three dee reconstruction featuring seven scenarios that reflect complex in-the-wild agricultural environments to illustrate the significant challenges involved <ref:2508.18694#pg1,complex in-the-wild agricultural>.
Taro: This benchmark specifically highlights the crucial role that the AgriChrono dataset plays in advancing research because it shows where current models struggle when faced with these real world conditions.
Conclusion: Rosa: So looking at the AgriChrono paper, with its focus on longitudinal data and novel view synthesis, the implication is that we finally have a way to move beyond just indoor lab testing for three dee crop reconstruction <ref:2508.18694#pg1>.
Dev: It moves us toward having robust AI models that can handle dynamic environments in agriculture because this dataset and benchmark suite force researchers to test their methods under real world complexities like lighting shifts and growth stages.
Taro: Essentially, the paper shows that you need these types of large scale, time synchronized datasets and tough benchmarks to make sure any new reconstruction method is actually reliable when applied outside a controlled setting.
Rosa: The authors developed this modular platform to tackle the challenge of building three dee digital twins for expansive agricultural environments by capturing continuous growth in the field <ref:2508.18694#pg1>.
Dev: The work suggests that for agriculture automation to become widespread, we need these kinds of comprehensive data sources and standardized ways to evaluate performance across diverse real world conditions.
Jaehwan Jeong, Tuan-Anh Vu, Mohammad Jony, Shahab Ahmad, Md. Mukhlesur Rahman, Sangpil Kim, †M. Khalid Jawed
Korea University · University of California, Los Angeles · North Dakota State University
cs.RO, cs.AI, cs.SY, eess.SY
Submitted: 2025-08-26
Updated: 2026-11-14
Comments: Keywords: Agricultural robotics, Multimodal dataset, Longitudinal crop monitoring, Illumination variability, 3D Gaussian Splatting, Novel view synthesis, 3D Reconstruction
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 82/100
The gist: Accurate 3D reconstruction of crop structures is essential for geometryaware agricultural automation, yet remains challenging due to occlusion, thin plant topology, and dynamic motion.
Key concepts
- AgriChrono
- A modular robotic data-collection platform integrating multiple sensors like RGB, Depth, LiDAR, IMU, and Pose data. It is designed for remote, time-synchronized acquisition of dynamic farmland conditions to build a large dataset.
- Multi-modal Dataset
- A massive collection of diverse sensor data captured over time (18 TB) documenting the entire growth cycle of canola under different illumination. This dataset captures the complex, 'in-the-wild' environment necessary for robust 3D crop reconstruction.
- Temporal Alignment Mechanism
- A central software trigger that writes to a shared sync_time.txt file to ensure precise temporal alignment between independent sensors (RGB, Depth, LiDAR). This guarantees that all data streams are accurately synchronized to a single, defined start time.
- Benchmark Scenarios
- Seven distinct testing conditions used in the new benchmark designed to isolate specific environmental stressors like lighting variability and morphological changes in the crop. This tests how well 3D reconstruction methods perform under diverse real-world complexities.
Terminology
Summary
Accurate 3D reconstruction of crop structures is essential for geometryaware agricultural automation, yet remains challenging due to occlusion, thin plant topology, and dynamic motion. p. 0 Khalid Jawed2,†
1Korea University 2University of California, Los Angeles 3North Dakota State University
Abstract. Accurate 3D reconstruction of crop structures is essential for geometryaware agricultural automation, yet remains challenging due to occlusion, thin
plant topology, and dynamic motion. This deficiency fundamentally
limits the robustness of AI models, particularly for scene-level dynamic 3D reconstruction in dynamic environments. In this paper, we present AgriChrono, a
modular robotic data-collection platform and a multi-modal dataset designed to
capture dynamic farmland conditions.Where the paper says it
Where the paper says it
Dataset Construction and Scale
The AgriChrono platform integrates multiple sensors, enabling remote, time-synchronized acquisition of RGB, Depth, LiDAR, IMU, and Pose data for efficient and repeatable long-term data collection in real-world agricultural environments. The researchers successfully collected 18 TB over one month documenting the entire growth cycle of Canola under diverse illumination conditions. p. 0 Our platform integrates multiple sensors,
enabling remote, time-synchronized acquisition of RGB, Depth, LiDAR, IMU,
and Pose data for efficient and repeatable long-term data collection in real-world
agricultural environments. We successfully collected 18 TB over one month, documenting the entire growth cycle of Canola under diverse illumination conditions. – We release the 18 TB AgriChrono dataset, which captures comprehensively the inthe-wild environment. The data were captured four times per day (under different
illuminations) during the crop’s main growth period, documenting continuous morphological changes in a complex natural setting.Where the paper says it
Where the paper says it
System Architecture and Data Synchronization
The platform is built on the AgileX Scout 2.0 Unmanned Ground Vehicle (UGV), integrating a tiered sensor suite, an onboard compute unit, and a wireless communication module. p. 5 3, the system is built on the AgileX Scout 2. 0 Unmanned Ground Vehicle (UGV),
integrating a tiered sensor suite, an onboard compute unit, and a wireless communication module. Sensor synchronization is
ensured through a shared sync_time. txt file generated at the start of each session,
guaranteeing consistent temporal alignment across all modalities. Sensor synchronization is
ensured through a shared sync_time. txt file generated at the start of each session,
guaranteeing consistent temporal alignment across all modalities.Where the paper says it
Where the paper says it
Where the paper says it
Benchmark Design and Evaluation Metrics
A new benchmark for 3D reconstruction is introduced featuring seven scenarios that isolate specific environmental stressors, such as lighting variability and morphological changes. The evaluation of state-of-the-art methods reveals significant performance variance across these scenarios, highlighting the lack of uniform robustness in existing models when faced with diverse real-world complexities. p. 1 Building on this high-fidelity data, we introduce a new 3D reconstruction benchmark for non-rigid, in-the-wild scenes, featuring seven scenarios that isolate specific
environmental stressors, such as lighting variability and morphological changes. Our
evaluation of state-of-the-art methods from the last three years reveals significant performance variance across these scenarios, highlighting the lack of uniform robustness
in existing models when faced with diverse real-world complexities. As shown in Table 4, the benchmark evaluates
two core challenges across seven scenarios. The first, Lighting Variance, consists of four
scenarios using Day 19 captures at four distinct times (6:00 AM, 11:00 AM, 4:00 PM,
9:00 PM) to assess robustness to lighting changes.Where the paper says it
Where the paper says it
Performance Analysis of Reconstruction Methods
The benchmark compares 16 state-of-the-art 3D reconstruction methods, including 12 Gaussian Splatting-based and 4 NeRF-based techniques across seven distinct conditions. Experimental results demonstrate that all baselines not only exhibit poor performance across scenarios but also fail to maintain consistent performance levels across them. p. 1 Building on this high-fidelity data, we introduce a new 3D reconstruction benchmark for non-rigid, in-the-wild scenes, featuring seven scenarios that isolate specific
environmental stressors, such as lighting variability and morphological changes. Our
evaluation of state-of-the-art methods from the last three years reveals significant performance variance across these scenarios, highlighting the lack of uniform robustness
in existing models when faced with diverse real-world complexities. 3DGS-based methods. The original Gaussian Splatting method [12] represents scenes
as anisotropic 3D Gaussians rendered via screen-space splatting, achieving NeRF-level
image quality with real-time rendering and fast training, and serves as our main explicit baseline. Scaling these capabilities
to complex real-world agricultural settings necessitates robust reconstruction performance as a foundational prerequisite. For instance, beyond thin-canopy challenges,
models must account for dynamic environmental factors, such as illumination shifts,
wind-induced motion, and non-rigid morphological changes, throughout the growth cycle to transition from the laboratory to unpredictable real-world farmland.Where the paper says it
Where the paper says it
Where the paper says it
Limitations and Future Directions
The data collection faced challenges, with Site 2 exhibiting lower FPS values due to thermal issues during prolonged operation. Immediate future work will involve refining the platform structure to mitigate thermal issues and developing software solutions for dynamic time synchronization using RTK-GNSS. The research efforts are directed toward scaling 3D Digital Twins from controlled indoor settings to expansive, in-the-wild environments. p. 0 Accurate 3D reconstruction of crop structures is essential for geometryaware agricultural automation, yet remains challenging due to occlusion, thin
plant topology, and dynamic motion. While controlled indoor environments facilitate high-quality object-level 3D reconstruction, scaling these methods to complex "in-the-wild" agricultural settings remains a formidable challenge due to
scarcity of dedicated datasets and benchmarks.Where the paper says it
Acknowledgements
This work was supported in part by the US Department of Agriculture (grant numbers 2024-67021-42528 and 2022-67022-37021), the Culture, Sports, and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2024 (International Collaborative Research and Global Talent Development for the Development of Copyright Management and Protection Technologies for Generative AI, RS-2024-00345025), and the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2019-II190079, Artificial Intelligence Graduate School Program, Korea University). p. 0 Mukhlesur Rahman3
, Sangpil Kim1,†
, and M. Khalid Jawed2,†
1Korea University 2University of California, Los Angeles 3North Dakota State University
Abstract.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Develop robust 3D reconstruction models capable of handling dynamic perturbations by training them on AgriChrono, which includes
continuous structural changes
andcolor shifts caused by the changing position of the sun.
This allows AI models to generalize beyond controlled indoor settings tocomplex 'in-the-wild' agricultural settings.
-
Enhance model performance under illumination variability by using the AgriChrono benchmark, which features scenarios testing four distinct lighting conditions. This enables AI systems to maintain high fidelity even when facing
intense, directional illumination
versus softer lighting. -
Improve non-rigid scene understanding by training methods on the AgriChrono dataset, which documents
continuous morphological changes in a complex natural setting.
This directly addresses the challenge of reconstructingnon-rigid, in-the-wild scenes.
-
Increase pose estimation reliability by leveraging the multi-modal nature of AgriChrono data to derive Visual-Inertial Odometry (VIO) poses with
consistent, low-variance inter-frame displacements,
leading tohighly robust robot trajectories.
-
Create a modular robotic data collection system that enables remote, hands-off operation via a customized WebUI. This platform allows for the
precise, time-synchronized acquisition of 4-view RGB, 2-view Depth, LiDAR, IMU, and VIO Pose
over long distances.
Abstract
Reliable 3D reconstruction of field-grown crops is hindered by thin, overlapping structures, wind-driven foliage motion, and changing illumination. Yet longitudinal field datasets and standardized novel-view synthesis (NVS) benchmarks remain scarce. We present an integrated framework for robotic acquisition, dataset construction, NVS evaluation, and agricultural scene reconstruction. AgriChrobot enables web-based teleoperation and repeated multi-view capture along crop rows. Using this platform, we introduce AgriChrono, a longitudinal canola dataset collected over approximately one month, comprising 160 recording sessions, four RGB streams, two depth maps, and estimated camera poses. We define seven NVS scenarios spanning four times of day and three acquisition dates, and benchmark 15 NeRF- and Gaussian Splatting-based methods. In addition to whole-image metrics, we assess fine-detail reconstruction within manually annotated canola regions using gradient- and edge-based metrics. Our results show that strong whole-image scores do not guarantee faithful reconstruction of thin, overlapping crop structures. To address this gap, we propose AgriSplat, which relaxes visibility filtering during initialization and reallocates model capacity through contribution-based pruning and budgeted reinvestment in sub-threshold candidates ranked by existing gradient statistics. After reallocation, asymmetric directional supervision penalizes insufficient contrast along reference image-gradient directions. Relative to Octree-GS, AgriSplat improves mean crop-region gradient-energy ratio from 0.506 to 0.847 and edge F-score from 0.712 to 0.797 across seven scenarios, while also improving whole-image PSNR, SSIM, and LPIPS. These contributions provide a dataset, benchmark, and method for systematic crop reconstruction under field conditions.
Sources
- Reconstruction Using the Invisible: Intuition from NIR and Metadata for Enhanced 3D Gaussian Splatting
- A survey of datasets for computer vision in agriculture
- Vision-Guided Targeted Grasping and Vibration for Robotic Pollination in Controlled Environments
- Your Vision-Language-Action Model Already Has Attention Heads For Path Deviation Detection
- Octree-GS: Towards Consistent Real-time Rendering with LOD-Structured 3D Gaussians
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation