Join us for a series of virtual events to hear talks from experts on the latest developments at the intersection of Visual AI in Agriculture.

Training robust perception models for robotics and autonomy often requires massive, diverse 3D datasets. But collecting and annotating real-world LiDAR point clouds at scale is both expensive and time-consuming, especially when high-quality labels are needed. Paved2Paradise introduces a cost-effective alternative: a scalable LiDAR simulation pipeline that generates realistic, fully annotated datasets with minimal human labeling effort.

The key idea is to “factor the real world” by separately capturing background scans (e.g., fields, roads, construction sites) and object scans (e.g., vehicles, people, machinery). By intelligently combining these two sources, Paved2Paradise can synthesize a combinatorially large set of diverse training scenes. The pipeline involves four steps: (1) collecting extensive background LiDAR scans, (2) recording high-resolution scans of target objects under controlled conditions, (3) inserting objects into backgrounds with physically consistent placement and occlusion, and (4) simulating LiDAR geometry to ensure realism.

Experiments show that models trained on Paved2Paradise-generated data transfer effectively to the real world, achieving strong detection performance with far less manual annotation compared to conventional dataset collection. The approach is not only cost-efficient, but also flexible—allowing practitioners to easily expand to new object classes or domains by swapping in new background or object scans.

For ML practitioners working in robotics, autonomous vehicles, or safety-critical perception, Paved2Paradise highlights a practical path toward scaling training data without scaling costs. It bridges the gap between simulation and real-world performance, enabling faster iteration and more reliable deployment of perception models.

Dr. Andy Quitmeyer will talk about the design of an exciting new open source science tool, The Mothbox. The Mothbox is an award winning project for broad scale monitoring of insects for biodiversity. It's a low cost device developed in harsh Panamanian jungles which takes super high resolution photos to then automatically ID the levels of biodiversity in forests and agriculture. After thousands of insect observations and hundreds of deployments in Panama, Peru, Mexico, Ecuador, and the US, we are now developing a new, manufacturable version to share this important tool worldwide. We will discuss the development of this device in the jungles of Panama and its importance to studying biodiversity worldwide.

Anomaly detection is transforming manufacturing and surveillance, but what about agriculture? Can AI actually detect plant diseases and pest damage early enough to make a difference?

This talk demonstrates how anomaly detection identifies and localizes crop problems using coffee leaf health as our primary example. We'll start with the foundational theory, then examine how these models detect rust and miner damage in leaf imagery.

The session includes a comprehensive hands-on workflow using the open-source FiftyOne computer vision toolkit, covering dataset curation, patch extraction, model training, and result visualization. You'll gain both theoretical understanding of anomaly detection in computer vision and practical experience applying these techniques to agricultural challenges and other domains.

Foundation models have enabled a new way to address tasks, by benefitting from emerging capabilities in a zero-shot manner. In this talk I will discuss recent research on enabling visual AI in a zero-shot manner and via fine-tuning. Specifically, I will discuss joint work on RELOCATE, a simple training-free baseline designed to perform the challenging task of visual query localization in long videos.

To eliminate the need for task-specific training and efficiently handle long videos, RELOCATE leverages a region-based representation derived from pretrained vision models. I will also discuss joint work on enabling multi-modal large language models (MLLMs) to correctly answer prompts that require a holistic spatio-temporal understanding: MLLMs struggle to answer prompts that refer to 1) the entirety of an environment that an agent equipped with an MLLM can operate in; and simultaneously also refer to 2) recent actions that just happened and are encoded in a video clip.

However, such a holistic spatio-temporal understanding is important for agents operating in the real world. Our solution involves development of a dedicated data collection pipeline and fine-tuning of an MLLM equipped with projectors to improve both spatial understanding of an environment and temporal understanding of recent observations.