Releases·Tool Release
Mozilla Data Collective releases 425-image mirror-suit dataset for stress-testing CV and depth-estimation against specular reflection
A 425-image dataset of a person in a custom faceted mirror suit was released via the Mozilla Data Collective to stress-test computer vision, depth-estimation, and spatial AI models against severe specular glare and geometric reflections. It ships as uncompressed Camera-Master RAWs plus high-resolution JPEGs with SHA-256 forensic manifests. Practitioners can use it to reproduce real-world failure modes like bounding-box dropouts and segmentation breakdowns in reflective scenes.
3 engagement1 sourceSun, Oct 4, 2026, 05:21 AM
AI Overview
- Mozilla Data Collective released a 425-image dataset of a faceted mirror suit shot outdoors to trigger bounding-box dropouts and segmentation failures.
- The archive includes uncompressed Camera-Master RAWs, high-resolution JPEGs, and block-buffered SHA-256 forensic manifests.
Related
Social·Community
augmenting large datasets to have more edge case data for training [D]
A Reddit user seeks feedback on a data augmentation idea for autonomous driving training. The proposal adapts large labeled daytime datasets into rare conditions like night, fog, rain, and glare, keeping labels intact. Practitioners may find the approach relevant for addressing distribution gaps in camera-based perception training data.
A Reddit user seeks feedback on a data augmentation idea for autonomous driving training. The proposal adapts large labeled daytime datasets into rare conditions like night, fog, rain, and glare, keeping labels intact. Practitioners may find the approach relevant for addressing distribution gaps in camera-based perception training data.
7 engagement1 sourceredditFri, Sep 18, 2026, 11:24 AM
Developments·Benchmark
GBU-Palm: Multimodal Video Dataset and Benchmark for Palm Presentation Attack Detection
Researchers released GBU-Palm, a large-scale multimodal video dataset and benchmark for palm presentation attack detection (PAD). It contains 21,326 videos from 105 subjects and 210 palms across six acquisition environments, including bona fide, Print, and Replay presentations, with 6,310 synchronized RGB-NIR samples. The dataset addresses limitations of existing PAD datasets by providing systematic evaluation across environments, modalities, and attack types.
Researchers released GBU-Palm, a large-scale multimodal video dataset and benchmark for palm presentation attack detection (PAD). It contains 21,326 videos from 105 subjects and 210 palms across six acquisition environments, including bona fide, Print, and Replay presentations, with 6,310 synchronized RGB-NIR samples. The dataset addresses limitations of existing PAD datasets by providing systematic evaluation across environments, modalities, and attack types.
0 engagement1 sourcearxivFri, Aug 14, 2026, 03:31 PM
Developments·Paper
A Benchmark & Dataset for Detecting AI-Manipulated Visual Evidence in the Court System
Researchers introduced a benchmark and dataset for evidentiary image authentication in court contexts, where generative tools let non-experts fabricate or alter photographic evidence. The corpus of 1,505 items spans surveillance, dashcam, and consumer-photo imagery with structured metadata, and baselines show current manipulation detectors remain unreliable for evidentiary use.
Researchers introduced a benchmark and dataset for evidentiary image authentication in court contexts, where generative tools let non-experts fabricate or alter photographic evidence. The corpus of 1,505 items spans surveillance, dashcam, and consumer-photo imagery with structured metadata, and baselines show current manipulation detectors remain unreliable for evidentiary use.
0 engagement1 sourcearxivTue, Sep 29, 2026, 03:17 PM
Products·launched
RightWayUp - open-source 360-degree image rotation detection model
RightWayUp is an open-source image rotation detection model that estimates how far an image is rotated from upright across all 360 degrees. Built by ORTUS AI after existing models showed low accuracy and false positives on CCTV frames, it solves misaligned camera detection for video analytics. The team also reports a JPEG shortcut they found in a common rotation benchmark.
RightWayUp is an open-source image rotation detection model that estimates how far an image is rotated from upright across all 360 degrees. Built by ORTUS AI after existing models showed low accuracy and false positives on CCTV frames, it solves misaligned camera detection for video analytics. The team also reports a JPEG shortcut they found in a common rotation benchmark.
1 engagement1 sourceredditWed, Sep 30, 2026, 02:42 PM
Developments·Paper
Study proposes synthetic benchmarks to fully replace real faces in face recognition evaluation
A new arXiv paper tests 12 synthetic face datasets against 7 real benchmarks, finding that synthetic data can now rival real data for training face recognition models. The authors argue that evaluation should also move to synthetic benchmarks to eliminate privacy and legal concerns entirely.
A new arXiv paper tests 12 synthetic face datasets against 7 real benchmarks, finding that synthetic data can now rival real data for training face recognition models. The authors argue that evaluation should also move to synthetic benchmarks to eliminate privacy and legal concerns entirely.
0 engagement1 sourcearxivThu, Jul 16, 2026, 12:47 PM