Datature

Day 3 at #MODEX and we've talked to more ops, automation, and quality leaders than we expected for the whole show 🤖 The demos that keep getting requested: torn packaging detection on the conveyor, pallet load verification at the dock, VLM-driven process monitoring on the assembly floor → all running on real customer footage, not staged setups. One last day to catch us. The fastest way to get a focused 1:1 → Ping Brandon Neustadter or Max Wasilko to book a dedicated demo slot!

Datature is at #MODEX next week 🤖 We will be running live demos of Datature's Vision AI on real warehouse footages → from catching torn and crushed packaging on the conveyor, to verifying pallet loading configuration before they leave the dock. Be sure to stop by and chat with our engineers on how teams can leverage Vision-Language Models to automate their systems, with surprisingly little data and setup.  ▎Booth A7024, Hall A  ▎April 13-16  ▎Georgia World Congress Center, Atlanta  ▎Brandon Neustadter Max Wasilko #MODEX2026 #VisionAI #WarehouseAutomation #ManufacturingAI

Training a defect detection model usually means labeling hundreds of broken parts. But in manufacturing, defects are rare. Some failure modes haven't even happened yet. You can't label what doesn't exist. Anomaly detection flips the script. Instead of teaching a model what "broken" looks like, you teach it what "normal" looks like. Anything outside that boundary gets flagged. We tested three approaches in Anomalib on the MVTec AD benchmark ↘ PaDiM, PatchCore, EfficientAd All three hit pretty good image-level detection on the bottle category. The differences showed up in pixel-level precision + training time. The best part? No bounding boxes, no polygons, no annotation files. Your "labeling" is sorting /good and /bad folders. What's the drawback of using this library? Read our full walkthrough with code, results, and analysis on the blog to find out - Link in comments 👇

New Tutorial - Fine-Tune Qwen3-VL on Your Own Dataset 🩻

Most CV engineers stop at imread and imshow. But production computer vision runs on a different set of functions entirely. We broke down the 7 OpenCV functions that show up again and again in real-world pipelines:  🔹 dnn.readNet - run neural network inference without framework dependencies  🔹 warpPerspective - fix skewed documents and misaligned parts  🔹 calcOpticalFlowPyrLK - track motion between video frames  🔹 createBackgroundSubtractorMOG2 - separate moving objects from static scenes  🔹 findContours - detect shapes, count parts, measure defects  🔹 Canny - edge detection (1986 algorithm, still the standard)  🔹 inRange + morphologyEx - segment objects by color in any lighting Each one includes runnable Python code, parameter guidance, and real examples - from document scanning to factory inspection to traffic analysis. We will be going deeper into how chaining various function can lead to transformative data pipelines in our next post. The best part: these functions chain together. Canny feeds contours. Background subtraction enables counting. OpenCV handles the image plumbing so your ML models can focus on the hard problems. Full Guide with Code 👇 https://lnkd.in/gxixyz3E

We’ve been talking to several infrastructure drone and inspection infrastructure companies about how they’re deploying AI for solar panel inspection - and the takeaway is clear: speed matters, but workflow integration matters more - models that generate panel-level, GPS-tagged tickets are key. In this comprehensive article, we break down the end-to-end playbook: why manual inspection doesn’t scale, the full solar defect taxonomy (hotspots, PID, diode failures, cracks, soiling, delamination), which imaging modalities catch what (RGB vs thermal IR vs EL/UVF/PL) We present the modern drone workflow from capture → orthomosaic → panel segmentation → defect detection → prioritized maintenance reports - ultimately showing how Datature has all the features to support this workflow from end to end. Read The Article → https://lnkd.in/grKwWD7t

Great updates by Vladimir Iglovikov and team over at Albumentations 🤩

Pose estimation has quietly become production-grade. Object detection tells you “a person is here.” Pose estimation tells you “left arm up, right knee bent” - by predicting keypoints (wrists, elbows, knees, ankles, etc.) and connecting them into a skeleton. That structural signal is why pose is showing up everywhere from sports analytics and rehab to safety monitoring and gesture interfaces. In our latest guide, we cover the parts that matter when you actually ship. If you’re building anything that needs posture, motion, or intent - this is the quick primer. Read The Blog → https://lnkd.in/gc3f78Pt

Vision AI and Agri-Robotics are way past pilots as of 2026, but production deployments of vision capabilities stall for one reason: generic pre-trained models don’t hold up in the field (domain shift, lighting/season drift, tiny targets). Teams often see ~20–40% accuracy drops moving from benchmarks to real farm imagery. When speaking to our users, this accuracy has to be above ~88% on average to justify moving past pilots. The practical playbook for agri-robotics PMs/devs: fine-tune on 500+ labeled field images, then quantize and run on-edge (connectivity is unreliable and cloud latency misses actuation timing). Ship a tight loop: deploy → collect edge cases → retrain on fresh field data each season, per unique customer. Datature's platform compresses that pipeline (label → train → evaluate → export → deploy via MQTT) so you can get to field-ready models in weeks, not months. Read Our Findings → https://lnkd.in/gVbAzaFM