RF-DETR drops real-time segmentation models that run in 3.4ms
Researchers just released RF-DETR segmentation under an Apache 2.0 license, introducing a new real-time vision baseline. It arrives in six model sizes, ranging from a Nano variant clocking 3.4 ms per image (40.3 mAP) to a 2XLarge variant reaching 49.9 mAP at 21.8 ms. Why it matters: Developers get an unconstrained, ultra-fast segmentation backbone built for edge devices and latency-critical vision tasks.
RF-DETR segmentation is a new real-time instance segmentation family released under a fully permissive Apache 2.0 license.
Instead of offering just a single checkpoint, the project ships six different model sizes right out of the box. On the lightweight end, the Nano variant clocks in at just 3.4 ms per image while delivering 40.3 mAP. On the heavier end, the 2XLarge variant pushes accuracy to 49.9 mAP at 21.8 ms per image.
What it brings to the table
Basically, RF-DETR compresses DETR-based architectures down to genuine real-time latency budgets. Developers get a continuous dial between edge-tier velocity and high-precision polygon extraction.
Whether you are wiring up robotics perception, drone navigation, or live video monitoring, there is an off-the-shelf footprint ready to fit tight compute constraints without manual pruning tricks.
Why it matters
We think the big draw here is pairing top-tier latency figures with an unrestricted Apache 2.0 license. A lot of cutting-edge vision backbones remain locked behind restrictive research-only terms, making commercial production rollouts a legal headache.
If your pipeline depends on low-latency video streaming or embedded hardware, RF-DETR provides an accessible, drop-in vision backbone worth benchmarking immediately.
Sources (1)
- @skalskip92(opens in a new tab)Originalx.com