A detailed project write-up on optimising YOLOv8 for fall detection across multiple scenarios using edge computing, covering deployment, inference optimisation, and performance tuning.

For developers hunting for the next generation of computer vision tools, Modelers.cn—Huawei’s Ascend AI ecosystem platform—has just dropped a potentially significant update. The latest addition to…

Real-time and accurate detection of crop pests in field environments is critical for intelligent plant protection and ensuring food security. However, practical farmland settings introduce substantial...

A deep learning model elevated non-specialist x-ray interpretation to expert levels for detecting hip bone damage.

Vision-language models (VLMs) can solve visual CAPTCHAs without prior training, but they perform each instance as a new problem, lacking improvement over time. Specialized detectors offer speed but are limited to trained categories and also do not learn from experience. This research investigates a system that improves with repeated use by combining a fine-tuned YOLOv8 detector with an open-weight VLM, routing tasks based on confidence. The system operates solely through screenshots and OS inputs, achieving an overall accuracy of 85.4% across 16 CAPTCHA classes. Crucially, every VLM answer generates a training label for the detector, allowing it to learn new categories quickly without human intervention. This self-training mechanism also enables self-repair. If a CAPTCHA operator perturbs images to disable the detector, the VLM remains unaffected and provides labels that help the detector recover. A simulated year-long arms race demonstrated the solver's consistent recovery against monthly CAPTCHA perturbations. A 70%-accurate open-weight VLM proved as effective as a perfect oracle in hardening the detector. This suggests that CAPTCHA defenses under-estimate the rapid adaptability of modern solvers.

A research team has enhanced YOLOv8-Pose with a StarNet backbone, motion-relevant attention, and a composite keypoint loss to jointly classify tennis actions and estimate 17 body keypoints from single

Vision-language models (VLMs) can solve visual CAPTCHAs without prior training, but they perform each instance as a new problem, lacking improvement over time. Specialized detectors offer speed but are limited to trained categories and also do not learn from experience. This research investigates a system that improves with repeated use by combining a fine-tuned YOLOv8 detector with an open-weight VLM, routing tasks based on confidence. The system operates solely through screenshots and OS inputs, achieving an overall accuracy of 85.4% across 16 CAPTCHA classes. Crucially, every VLM answer generates a training label for the detector, allowing it to learn new categories quickly without human intervention. This self-training mechanism also enables self-repair. If a CAPTCHA operator perturbs images to disable the detector, the VLM remains unaffected and provides labels that help the detector recover. A simulated year-long arms race demonstrated the solver's consistent recovery against monthly CAPTCHA perturbations. A 70%-accurate open-weight VLM proved as effective as a perfect oracle in hardening the detector. This suggests that CAPTCHA defenses under-estimate the rapid adaptability of modern solvers.

Object detection becomes truly powerful when the model understands your world instead of just COCO-style benchmarks. That’s where yolov8…

YouTube video by OpenCV

A YOLOv8-based deep learning model identified and measured children's coronary arteries on echocardiograms with expert-level accuracy in under two seconds per measurement.

Hi everyone, I’m an independent researcher preparing my first arXiv submission and looking for an endorser in **cs.CV** (cross-listed to cs.AR and eess.IV). **Paper:** _Low-Latency Close-Range Drone Detection with Quantized YOLOv8 on a Zynq UltraScale+ MPSoC_ **PDF:** drone-detect/paper/main.pdf at master · beatwad/drone-detect · GitHub Code, weights and build recipe are in the same repo. **In short:** a vision chain for an aim-assistance system that detects drones at up to 10 m and reports their offset from the boresight. * **Model:** a single-head, stride-8 YOLOv8n with ReLU6, trained on 36,903 images merged from nine public datasets, with cluster-safe deduplication and per-distance-regime evaluation. * **Quantization:** W4A4 quantization-aware training (QAT) with Brevitas keeps mAP50 at 0.984 (close range) and 0.988 (mid range), with box-centre error unchanged from the float model. * **Hardware:** compiled with FINN into an FPGA dataflow accelerator on a ZCU102, running at 168 FPS with 22.6 ms latency, bit-exact against simulation. * **Live system:** with a USB camera, the full camera → detector → tracker chain runs at ~113 FPS with a 35.8 ms median latency. * **Limitations:** the paper also covers what doesn’t work yet, including false alarms on birds and aircraft and no evaluation on real footage through the deployment lens. If you have recent cs.CV papers on arXiv and are willing to take a look, the endorsement code is **LEZ4O6** , or you can endorse directly at Log in to arXiv | arXiv e-print repository . arXiv only asks endorsers to check that the paper is appropriate for the category, not to review it in detail. I’m happy to answer any questions about the work. Thank you!

Researchers have built a deep eutectic fluorescent composite paired with YOLOv8 computer vision that detects mercury ions at 60 nanomolar sensitivity and displays traffic-light-style safety verdicts

YOLOv8 runs on the NPU of the GK7205V5x0 with the models in flash and no setup, a pan/tilt camera follows a person and stays sharp through the move, and anti-flicker stops guessing and tests the light...

Phishing remains one of the most common vectors for financial and identity fraud, and most detection systems still rely on inspecting a page's URL, HTML markup, or domain registration history. These signals are easy for an attacker to rotate or obfuscate, and they say very little about what actuall…

Phishing remains one of the most common vectors for financial and identity fraud, and most detection systems still rely on inspecting a page's URL, HTML markup, or domain registration history. These signals are easy for an attacker to rotate or obfuscate, and they say very little about what actuall…

A 2025 study compares YOLOv8 across Raspberry Pi 5 and Jetson Orin hardware, revealing the trade-offs behind real-time drone vision.

When you train YOLOv8 on custom dataset, you turn a general-purpose object detector into a specialist that understands exactly the objects…

I maintain an open-source app that turns long streams into vertical clips entirely on the user’s machine. Sharing it here for one specific reason: the part that decides which moments matter is deliberately swappable, and I would rather other people fill it than guess at it myself. What runs today: YOLOv8 finds people, OpenCV Haar cascades find the face inside each person box, and TalkNet-ASD decides which face is actually speaking so the 16:9 to 9:16 crop follows the speaker. That last model earns its place. The intuitive approach is to follow whichever mouth is moving, and it does not work. On a real two-person clip, mouth-region pixel motion scored the speaker 0.0746 and the listener 0.0801, a 7% difference in the wrong direction, while their on-screen sizes differed by 45%. Pixel motion measures movement, not speech, so anything built on it follows whoever is nearest the camera. The open slot is the scoring side. Which moments become clips is decided by per-second signals that get percentile-ranked and scanned for peaks. A detector that emits a confidence over the timeline becomes one more signal and its peaks become candidate clips, with nothing in the scorer changing. So if you have a model for footage this was never tuned for, sports, gaming, reactions, lectures, wildlife, it can drive clip selection without rewriting the pipeline. I am not building any of those. I would rather the seam exist and other people use it. Contract, the file to copy and the performance trap: github.com/ColinGPT9/clips-studio #### docs/EXTENDING.md main # Extending Clips Kitty The changes people most often want to make, and what each actually touches. All of them are deliberately small, if one of these turns into a sprawling diff, something has drifted and the design is worth a second look. Read [ARCHITECTURE.md](../ARCHITECTURE.md) first for how the pieces fit. If you want to build something *beside* the app rather than change it. A bot, a batch runner, another front end. You probably want [API.md](API.md) instead. The whole pipeline is already reachable over HTTP. --- ## Add a language Translation, subtitles, dubbing and the interface all key off one table. 1. **`multilingual/languages.py`**: add a row to `LANGUAGES`: ```python This file has been truncated. show original AGPL-3.0, Windows, nothing leaves the machine.

Bringing object detection from YouTube video into focus