What BEAMR compression technology is?
Best-in-class video data compression
Beamr delivers ML-safe video data compression for AV and ADAS in 2 modes: Lossless bit-exact with 12-bit Bayer; and Content-Adaptive.
ML-Safe Lossless (coming soon): 12-bit Bayer in, 12-bit Bayer out – in real time. It uses the encoder already in the GPU that typically sits idle – while the GPU’s compute cores stay free for other workloads. In initial Beamr tests on 12-bit Bayer recordings from 8 cameras, 8-megapixel each, compressed files were 47% smaller than raw.
ML-Safe Content-Adaptive (lossy): Built on Beamr’s patented Content-Adaptive Bitrate technology (CABR). Up to 50% file size reduction while preserving the visual detail that perception and detection models rely on. Rather than applying fixed settings to every frame, CABR analyzes each frame and compresses it as far as the content allows – holding onto the object boundaries, structural edges, and scene geometry.
ML-Safe compression fits into your existing infrastructure and workflows, reducing file sizes without changing how downstream models behave. It is GPU-accelerated for high throughput, ingests any input format, and outputs standard codecs – AVC, HEVC, AV1 – so compressed data stays compatible with your existing tools.
How it fits in RTMaps pipelines
- In-vehicle capture and logging Bring ML-Safe compression - Lossless or Content-Adaptive - into the logging pipeline so constrained in-car storage holds more recorded video.
- Cloud and data center Batch-compress existing recorded archives to cut petabyte-scale storage and egress on data you already hold with ML-safe results.
- Simulation and model training Compress the real-world and synthetic datasets feeding sensor- and software-in-the-loop testing and training, including world-foundation-model output, so less data moves through training and validation.
beamr gallery
Benchmarked across the Autonomous vehicle pipeline
Beamr tests ML-safe CABR on real-world AV footage and industry-standard datasets, measuring model-level metrics rather than pixel quality alone. Benchmarking has demonstrated up to 50% file size reduction with no measurable impact on model accuracy.
- In the RTMaps pipeline: on real-world sequences processed through RTMaps, CABR cut file size 97% versus uncompressed and 31% versus baseline encodes, while preserving model accuracy. (Learn more)
- Object detection: roughly 48% smaller files, under 2% mAP difference, minimal localization shift, and confidence scores that stay highly correlated with the uncompressed baseline. (Learn more)
- World foundation model, NVIDIA Cosmos Curator - Captioning and scene curation: 41–57% improved compression with no measurable impact on the captioning pipeline. Agreement rates between variants remained between 88.9% and 100% across all tested videos, with most videos achieving 93%–98% agreement. (Learn more)
- Model robustness: a depth-estimation model (Depth Anything V2) trained on CABR-compressed footage showed 30.7% lower depth error on vulnerable road users (VRUs) and 16% lower across all object classes, at 35.2% smaller files - first evidence that compressed data can strengthen model resilience, not just preserve it. (Learn more)
BENEFITS AND KEY ADVANTAGES OF THE ML-SAFE VIDEO DATA COMPRESSION BY BEAMR
- ML-safe compression with up to 50% smaller files - Lossless: bit-exact, 12-bit Bayer (coming soon); Content-adaptive, benchmarked on real-world and synthetic AV video data, with under 2% mAP difference, minimal localization shift, and confidence scores that stay highly correlated.
- Lower storage, egress, and I/O costs, faster data transfer, and freeing resources for development - at petabyte scale, on data you already have.
- Fits your stack - runs as an RTMaps component across in-car, simulation, and cloud; GPU-accelerated; outputs AVC, HEVC, AV1.
- Processed in your own infrastructure - nothing leaves your environment, full privacy.
ABOUT BEAMR
Beamr (NASDAQ: BMR) delivers patented, proven video data compression technology for AV and ADAS, with partners including NVIDIA and AWS. Benchmarked on industry-standard datasets, and engaged with AV programs to address the petabyte-scale video data challenges in AV and Physical AI development.
For more informations, contact Beamr and Intempora
Please, fill the form below.