Berlin AI Machine Learning and Computer Vision Meetup

Location
Virtual. Register for the Zoom.

Field-Ready Vision: Building the Agricultural Image Repository (AgIR) for Sustainable Farming

Data—not models—is the bottleneck in agricultural computer vision. This talk shares how Precision Sustainable Agriculture (PSA) is tackling that gap with the Agricultural Image Repository (AgIR): a cloud bank of high-resolution, labeled images spanning weeds (40+ species), cover crops, and cash crops across regions, seasons, and sensors.

We’ll show how AgIR blends two complementary streams:

(1) semi-field, high-throughput data captured by BenchBot, our open-source, modular gantry that autonomously images plants and feeds a semi-automated annotation pipeline;
(2) true field images that capture real environmental variability. Together, they cut labeling cost, accelerate pretraining, and improve robustness in production.

On top of AgIR, we’ve built a data-centric training stack: hierarchical augmentation groups, batch mixers, a stand-alone visualizer for rapid iteration, and a reproducible PyTorch Lightning pipeline. We’ll cover practical lessons from segmentation (crop/weed/residue/water/soil), handling domain shift between semi-field and field scenes, and designing metadata schemas that actually pay off at model time.

About the Speaker

Sina Baghbanijam is a Ph.D. candidate in Electrical and Computer Engineering at North Carolina State University, where his research centers on generative AI, computer vision, and machine learning. His work bridges advanced AI methods with real-world applications across agriculture, medicine, and the social sciences, with a focus on large-scale image segmentation, bias-aware modeling, and data-driven analysis. In addition to his academic research, Sina is currently serving as an Agricultural Image Repository Software Engineering Intern with Precision Sustainable Agriculture, where he develops scalable pipelines and metadata systems to support AI-driven analysis of crop, soil, and field imagery.

Beyond Manual Measurements: How AI is Accelerating Plant Breeding

Traditional plant breeding relies on manual phenotypic measurements that are time-intensive, subjective, and create bottlenecks in variety development. This presentation demonstrates how computer vision and artificial intelligence are revolutionizing plant selection processes by automating trait extraction from simple photographs. Our cloud-based platform transforms images captured with smartphones, drones, or laboratory cameras into instant, quantitative phenotypic data including fruit count, size measurements, and weight estimations.

The system integrates phenotypic data with genotypic, pedigree, and environmental information in a unified database, enabling real-time analytics and decision support through intuitive dashboards. Unlike expensive hardware-dependent solutions, our software-focused approach works with existing camera equipment and standard breeding workflows, making advanced phenotyping accessible to organizations of all sizes.

About the Speaker

Dr. Sharon Inch is a botanist with a PhD in Plant Pathology and over 20 years of experience in horticulture and agricultural research. Throughout her career, she has witnessed firsthand the inefficiencies of traditional breeding methods, inspiring her to found AgriVision Analytics. As CEO, she leads the development of cloud-based computer vision platforms that transform plant breeding workflows through AI-powered phenotyping. Her work focuses on accelerating variety development and improving breeding decision-making through automated trait extraction and data integration. Dr. Sharon Inch is passionate about bridging the gap between advanced technology and practical agricultural applications to address global food security challenges.

AI-assisted sweetpotato yield estimation pipelines using optical sensor data

In this presentation, we will introduce the sensor systems and AI-powered analysis algorithms used in high-throughput sweetpotato post-harvest packing pipelines (developed by the Optical Sensing Lab at NC State University). By collecting image data from sweetpotato fields and packing lines respectively, we aim to quantitatively optimize the grading and yield estimation process, and the planning on storage and inventory-order matching.

We built two customized sensor devices to collect data respectively from the top bins when receiving sweetpotatoes from farmers, and eliminator table before grading and packing process. We also developed a compact instance segmentation pipeline that can run on smart phones for rapid yield estimation in-field with resource limitations. To minimize data privacy concerns and Internet connectivity issues, we try to keep all the analysis pipelines on the edge, which results in a design tradeoff between resource availability and environmental constraints. We will also introduce sensor building with these considerations. The analysis results and real time production information are then integrated into an interactive online dashboard, where stakeholders can leverage to help with inventory-order management and making operational decisions.

About the Speaker

Yifan Wu is a current Ph.D candidate at NC State University working in the Optical Sensing Lab (OSL) supervised by Dr. Michael Kudenov. Research focuses on developing sensor systems and machine learning platforms for business intelligence applications.

An End-to-End AgTech Use Case in FiftyOne

The agricultural sector is increasingly turning to computer vision to tackle challenges in crop monitoring, pest detection, and yield optimization. Yet, developing robust models in this space often requires careful data exploration, curation, and evaluation—steps that are just as critical as model training itself.

In this talk, we will walk through an end-to-end AgTech use case using FiftyOne, an open-source tool for dataset visualization, curation, and model evaluation. Starting with a pest detection dataset, we will explore the samples and annotations to understand dataset quality and potential pitfalls. From there, we will curate the dataset by filtering, tagging, and identifying edge cases that could impact downstream performance. Next, we’ll train a computer vision model to detect different pest species and demonstrate how FiftyOne can be used to rigorously evaluate the results. Along the way, we’ll highlight how dataset-centric workflows can accelerate experimentation, improve model reliability, and surface actionable insights specific to agricultural applications.

By the end of the session, attendees will gain a practical understanding of how to:

- Explore and diagnose real-world agricultural datasets
- Curate training data for improved performance
- Train and evaluate pest detection models
- Use FiftyOne to close the loop between data and models

This talk will be valuable for anyone working at the intersection of agriculture and computer vision, whether you’re building production models or just beginning to explore AgTech use cases.

About the Speaker

Prerna Dhareshwar is a Machine Learning Engineer at Voxel51, where she helps customers leverage FiftyOne to accelerate dataset curation, model development, and evaluation in real-world AI workflows. She brings extensive experience building and deploying computer vision and machine learning systems across industries. Prior to Voxel51, Prerna was a Senior Machine Learning Engineer at Instrumental Inc., where she developed models for defect detection in manufacturing, and a Machine Learning Software Engineer at Pure Storage, focusing on predictive analytics and automation.

* Oct 28, 2025
* 9:00-10:30 AM Pacific
* Online. Register for the Zoom!

Want greater visibility into the quality of your computer vision datasets and models? Then join us for this free 90-minute, hands-on workshop to learn how to leverage the open source FiftyOne computer vision toolset.
At the end of the workshop, you’ll be able to:

• Load and visualize agricultural datasets with complex labels
• Explore data insights interactively using embeddings and statistics
• Work with geolocation and map-based visualizations
• Generate high-quality annotations with the Segment Anything Model (SAM2)
• Evaluate model performance and debug predictions using real AgTec scenarios

Prerequisites: working knowledge of Python and basic computer vision concepts.

Resources: All attendees will get access to the tutorials, videos, and code examples used in the workshop.

Learn how to:

• Visualize complex datasets
• Explore embeddings
• Analyze and improve models
• Perform advanced data curation
• Build integrations with popular ML tools, models, and datasets

Oct 30, 2025
9 AM Pacific
Online. Register for the Zoom!

The Agent Factory: Building a Platform for Enterprise-Wide AI Automation

In this talk we will explore what it takes to build an enterprise-ready AI automation platform at scale. The topics covered will include:

• The Scale Challenge: E-commerce environments expose the limitations of single-point AI solutions, which create fragmented ecosystems lacking cohesion and efficient resource sharing across complex, knowledge-based work.
• Root Cause Analysis Success: Flipkart’s initial AI agent transformed business analysis from days-long investigations to near-instantaneous insights, proving the concept while revealing broader platform opportunities.
• Platform Strategy Evolution: Success across Engineering (SDLC, SRE), Operations, and Commerce teams necessitated a unified, multi-tenant platform serving diverse use cases with consistency and operational efficiency.
• Architectural Foundation: Leveraging framework-agnostic design principles we were able to emphasize modularity, which enabled teams to leverage different AI models while maintaining consistent interfaces and scalable infrastructure.
• The “Agent Garden” Vision: Flipkart’s roadmap envisions an internal ecosystem where teams discover, deploy, and contribute AI agents, providing a practical blueprint for scalable AI agent infrastructure development.

About the Speaker

Virender Bhargav at Flipkart is a seasoned engineering leader whose expertise spans business technology integration, enterprise applications, system design/architecture, and building highly scalable systems. With a deep understanding of technology, he has spearheaded teams, modernized technology landscapes, and managed core platform layers and strategic products. With extensive experience driving innovation at companies like Paytm and Flipkart, his contributions have left a lasting impact on the industry.

Scaling Generative Models at Scale with Ray and PyTorch

Generative image models like Stable Diffusion have opened up exciting possibilities for personalization, creativity, and scalable deployment. However, fine-tuning them in production‐grade settings poses challenges: managing compute, hyperparameters, model size, data, and distributed coordination are nontrivial.

In this talk, we’ll dive deep into learning how to fine-tune Stable Diffusion models using Ray Train (with HuggingFace Diffusers), including approaches like DreamBooth and LoRA. We’ll cover what works (and what doesn’t) in scaling out training jobs, handling large data, optimizing for GPU memory and speed, and validating outputs. Attendees will come away with practical insights and patterns they can use to fine-tune generative models in their own work.

About the Speaker

Suman Debnath is a Technical Lead (ML) at Anyscale, where he focuses on distributed training, fine-tuning, and inference optimization at scale on the cloud. His work centers around building and optimizing end-to-end machine learning workflows powered by distributed computing framework like Ray, enabling scalable and efficient ML systems.
Suman’s expertise spans Natural Language Processing (NLP), Large Language Models (LLMs), and Retrieval-Augmented Generation (RAG).
Earlier in his career, he developed performance benchmarking and monitoring tools for distributed storage systems. Beyond engineering, Suman is an active community contributor, having spoken at over 100 global conferences and events, including PyCon, PyData, ODSC, AIE and numerous meetups worldwide.

Privacy-preserving in Computer Vision through Optics Learning

Cameras are now ubiquitous, powering computer vision systems that assist us in everyday tasks and critical settings such as operating rooms. Yet, their widespread use raises serious privacy concerns: traditional cameras are designed to capture high-resolution images, making it easy to identify sensitive attributes such as faces, nudity, or personal objects. Once acquired, such data can be misused if accessed by adversaries. Existing software-based privacy mechanisms, such as blurring or pixelation, often degrade task performance and leave vulnerabilities in the processing pipeline.

In this talk, we explore an alternative question: how can we preserve privacy before or during image acquisition? By revisiting the image formation model, we show how camera optics themselves can be learned and optimized to acquire images that are unintelligible to humans yet remain useful for downstream vision tasks like action recognition. We will discuss recent approaches to learning camera lenses that intentionally produce privacy-preserving images, blurry and unrecognizable to the human eye, but still effective for machine perception. This paradigm shift opens the door to a new generation of cameras that embed privacy directly into their hardware design.

About the Speaker

Carlos Hinojosa is a Postdoctoral researcher at King Abdullah University of Science and Technology (KAUST) working with Prof. Bernard Ghanem. His research interests span Computer Vision, Machine Learning, AI Safety, and AI for Science. He focuses on developing safe, accurate, and efficient vision systems and machine-learning models that can reliably perceive, understand, and act on information, while ensuring robustness, protecting privacy, and aligning with societal values.

It's a (Blind) Match! Towards Vision-Language Correspondence without Parallel Data

Can we match vision and language embeddings without any supervision? According to the platonic representation hypothesis, as model and dataset scales increase, distances between corresponding representations are becoming similar in both embedding spaces. Our study demonstrates that pairwise distances are often sufficient to enable unsupervised matching, allowing vision-language correspondences to be discovered without any parallel data.

About the Speaker

Dominik Schnaus is a third-year Ph.D. student in the Computer Vision Group at the Technical University of Munich (TUM), supervised by Daniel Cremers. His research centers on multimodal and self-supervised learning with a special emphasis on understanding similarities across embedding spaces of different modalities.

Nov 19, 2025
9 AM Pacific
Online. Register for the Zoom!

AnimalClue: Recognizing Animals by their Traces

Wildlife observation plays an important role in biodiversity conservation, necessitating robust methodologies for monitoring wildlife populations and interspecies interactions. Recent advances in computer vision have significantly contributed to automating fundamental wildlife observation tasks, such as animal detection and species identification. However, accurately identifying species from indirect evidence like footprints and feces remains relatively underexplored, despite its importance in contributing to wildlife monitoring.

To bridge this gap, we introduce AnimalClue, the first large-scale dataset for species identification from images of indirect evidence. Our dataset consists of 159,605 bounding boxes encompassing five categories of indirect clues: footprints, feces, eggs, bones, and feathers. It covers 968 species, 200 families, and 65 orders. Each image is annotated with species-level labels, bounding boxes or segmentation masks, and fine-grained trait information, including activity patterns and habitat preferences. Unlike existing datasets primarily focused on direct visual features (e.g., animal appearances), AnimalClue presents unique challenges for classification, detection, and instance segmentation tasks due to the need for recognizing more detailed and subtle visual features. In our experiments, we extensively evaluate representative vision models and identify key challenges in animal identification from their traces.

About the Speaker

Risa Shinoda received her M.S. and Ph.D. in Agricultural Science from Kyoto University in 2022 and 2025. Since April 2025, she has been serving as a Specially Appointed Assistant Professor at the Graduate School of Information Science and Technology, the University of Osaka. She is engaged in research on the application of image recognition to plants and animals, as well as vision-language models.

LOTS of Fashion! Multi-Conditioning for Image Generation via Sketch-Text Pairing

Fashion design is a complex creative process that blends visual and textual expressions. Designers convey ideas through sketches, which define spatial structure and design elements, and textual descriptions, capturing material, texture, and stylistic details. In this paper, we present LOcalized Text and Sketch for fashion image generation (LOTS), an approach for compositional sketch-text based generation of complete fashion outlooks. LOTS leverages a global description with paired localized sketch + text information for conditioning and introduces a novel step-based merging strategy for diffusion adaptation.

First, a Modularized Pair-Centric representation encodes sketches and text into a shared latent space while preserving independent localized features; then, a Diffusion Pair Guidance phase integrates both local and global conditioning via attention-based guidance within the diffusion model’s multi-step denoising process. To validate our method, we build on Fashionpedia to release Sketchy, the first fashion dataset where multiple text-sketch pairs are provided per image. Quantitative results show LOTS achieves state-of-the-art image generation performance on both global and localized metrics, while qualitative examples and a human evaluation study highlight its unprecedented level of design customization.

About the Speaker

Federico Girella is a third-year Ph.D. student at the University of Verona (Italy), supervised by Prof. Marco Cristani, with expected graduation in May 2026. His research involves joint representations in the Image and Language multi-modal domain, working with deep neural networks such as (Large) Vision and Language Models and Text-to-Image Generative Models. His main body of work focuses on Text-to-Image Retrieval and Generation in the Fashion domain.

ProtoMedX: Explainable Multi-Modal Prototype Learning for Bone Health Assessment

Early detection of osteoporosis and osteopenia is critical, yet most AI models for bone health rely solely on imaging and offer little transparency into their decisions. In this talk, I will present ProtoMedX, the first prototype-based framework that combines lumbar spine DEXA scans with patient clinical records to deliver accurate and inherently explainable predictions.

Unlike black-box deep networks, ProtoMedX classifies patients by comparing them to learned case-based prototypes, mirroring how clinicians reason in practice. Our method not only achieves state-of-the-art accuracy on a real NHS dataset of 4,160 patients but also provides clear, interpretable explanations aligned with the upcoming EU AI Act requirements for high-risk medical AI. Beyond bone health, this work illustrates how prototype learning can make multi-modal AI both powerful and transparent, offering a blueprint for other safety-critical domains.

About the Speaker

Alvaro Lopez is a PhD candidate in Explainable AI at Lancaster University and an AI Research Associate at J.P. Morgan in London. His research focuses on prototype-based learning, multi-modal AI, and AI security. He has led projects on medical AI, fraud detection, and adversarial robustness, with applications ranging from healthcare to financial systems.

CLASP: Adaptive Spectral Clustering for Unsupervised Per-Image Segmentation

We introduce CLASP (Clustering via Adaptive Spectral Processing), a lightweight framework for unsupervised image segmentation that operates without any labeled data or fine-tuning. CLASP first extracts per-patch features using a self-supervised ViT encoder (DINO); then, it builds an affinity matrix and applies spectral clustering. To avoid manual tuning, we select the segment count automatically with a eigengap-silhouette search, and we sharpen the boundaries with a fully connected DenseCRF. Despite its simplicity and training-free nature, CLASP attains competitive mIoU and pixel-accuracy on COCO-Stuff and ADE20K, matching recent unsupervised baselines. The zero-training design makes CLASP a strong, easily reproducible baseline for large unannotated corpora—especially common in digital advertising and marketing workflows such as brand-safety screening, creative asset curation, and social-media content moderation.

About the Speaker

Max Curie is a Research Scientist at Integral Ad Science, building fast, lightweight solutions for brand safety, multi-media classification, and recommendation systems. As a former nuclear physicist at Princeton University, he brings rigorous analytical thinking and modeling discipline from his physics background to advance ad tech.