Skip to content
Apixo
Blog
news· 3 min read· via Roboflow Blog

RF-DETR Atto Delivers 1 ms Object Detection on NVIDIA T4 GPUs

Roboflow benchmark details the compact RF-DETR model family, showing sub-2ms inference times and competitive precision against YOLO26-N on NVIDIA T4 GPUs.

RF-DETR Atto Delivers 1 ms Object Detection on NVIDIA T4 GPUs

Computer vision projects operating under tight latency constraints now have access to a set of compact vision models designed for sub-two-millisecond inference. Benchmarks released on the RF-DETR framework show that its smallest variant, RF-DETR Atto, achieves a median engine latency of 1.0 ms per image when running on an NVIDIA T4 GPU at batch size one. The compact model family—comprising Atto, Femto, and Pico—targets high-speed real-time applications such as counting livestock at watering stations, monitoring truck occupancy in loading bays, and tracking parked vehicles.

These lightweight detectors are optimized for resource-constrained deployments while offering distinct trade-offs between precision and processing time. Built on top of Meta's PE-Core-T vision encoder described by Daniel Bolya et al., the models take square input images and convert them into grid patches. An internal adapter merges feature maps from the 3rd, 6th, 9th, and 12th encoder layers to form a consolidated spatial representation used for proposal generation and classification.

Architectural Differences Across the Models

Despite sharing a common parameter scale between 7.4 million and 8.4 million parameters—significantly smaller than the 30.5 million parameters found in RF-DETR Nano—the three variants utilize different spatial resolutions, attention mechanisms, and decoding strategies configured through neural architecture search (based on work by Robinson et al., ICLR 2026).

RF-DETR Atto processes 380×380 pixel inputs using 20×20 patches on a 19×19 grid. It applies global attention across all 12 encoder layers without using a decoder, outputting 300 candidate queries directly to achieve maximum speed.

RF-DETR Femto increases resolution to 384×384 with 16×16 patches (24×24 grid) and uses a localized attention structure. It divides the grid into 2×2 local windows across two consecutive layers before executing a global attention layer, repeating this pattern four times. Femto includes two decoder layers and processes 200 queries.

RF-DETR Pico pushes input size to 560×560 pixels with 20×20 patches on a 28×28 grid. It adopts the same 2×2 window pattern as Femto but increases the decoder depth to three layers with 200 candidate queries.

Performance Benchmarks on NVIDIA T4

Evaluated under Single Artifact Benchmarking using TensorRT 10.4 in 16-bit floating point (FP16) on an NVIDIA T4 GPU, the models demonstrate clear scaling across speed and accuracy on the COCO dataset (mAP@50:95):

  • RF-DETR Atto: 1.0 ms latency, 7.4M parameters, 30.5 COCO mAP, 48.3 RF100-VL mAP.
  • RF-DETR Femto: 1.4 ms latency (1.377 ms engine time), 7.9M parameters, 37.8 COCO mAP, 53.8 RF100-VL mAP.
  • RF-DETR Pico: 1.7 ms latency, 8.4M parameters, 41.6 COCO mAP, 56.0 RF100-VL mAP.

When compared with competing real-time detectors, RF-DETR Pico outperforms YOLO26-N at the exact same latency of 1.7 ms, where YOLO26-N scored 40.3 COCO mAP with 2.6M parameters. Against the DEIMv2 benchmark family, RF-DETR models consistently deliver higher precision at similar speeds; for instance, DEIMv2 Femto achieved 31.0 mAP at 1.4 ms compared to RF-DETR Femto's 37.8 mAP.

The performance advantages extend to specialized fine-tuning tasks tested on the 100-dataset RF100-VL benchmark over 100 epochs. RF-DETR Femto reached 53.8 mAP and Pico achieved 56.0 mAP across test sets, compared to 48.3 mAP for Atto. DEIMv2 variants struggled on datasets with small target objects—such as aerial livestock photos—due to training logic that filtered out small ground-truth bounding boxes.

What it means for developers

For engineers building real-time vision pipelines, the RF-DETR lightweight family offers a practical deployment path via PyTorch and TensorRT FP16 compilation. Developers can select a variant based on their precise latency budget—spending an additional 0.4 ms for Femto yields a 7.3 point mAP gain over Atto, while Pico adds another 0.3 ms for a 3.8 point improvement.

When building complex multimodal systems alongside edge vision deployments, developers can try top AI models cheaply through one API at https://apixoai.online.

From a licensing standpoint, Atto, Femto, and Pico are distributed under the Roboflow Platform Model License 1.0. Commercial deployments hosted directly on Roboflow's platform are permitted under standard terms, while self-hosted commercial deployments require Roboflow's Enterprise add-on.


Source: How RF-DETR Atto Detects Objects in 1 ms on an NVIDIA T4 — Roboflow Blog. Written by the Apixo team from that report.

#ai-news#object-detection#rf-detr#computer-vision#nvidia-t4#benchmarks
Try it with your own tools

One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.

Get your API key

Keep reading