mudler/rfdetr-cpp-seg-2xlarge

RF-DETR Seg-2XLarge — GGUF for rfdetr.cpp

GGUF-format weights of Roboflow RF-DETR Seg-2XLarge (segmentation variant) for use with rfdetr.cpp, a C++/ggml implementation that matches the upstream PyTorch model on CPU.

This repo contains all four standard quantizations of this variant. F16 is the recommended default — near-identical accuracy to F32 (see the table below), 1.83× smaller than F32, and it takes ggml's F32×F16 matmul fast path.

Available files

File	Quant	Size (MB)	Recall @ IoU 0.5	Recall @ IoU 0.95	Mean mask IoU	Pixel agreement	Latency (median ms, T=8)
rfdetr-seg-2xlarge-f32.gguf	F32	143.9	1.0000	0.9870	0.9960	0.9999	—
rfdetr-seg-2xlarge-f16.gguf ← recommended	F16	78.7	1.0000	0.9870	0.9964	0.9999	—
rfdetr-seg-2xlarge-q8_0.gguf	Q8_0	48.2	1.0000	1.0000	0.9937	0.9999	—
rfdetr-seg-2xlarge-q4_K.gguf	Q4_K	38.6	0.9417	0.7102	0.9732	0.9994	—

All accuracy numbers are computed against the upstream PyTorch reference (rfdetr 1.9.0) on 7 COCO val2017 images at threshold 0.5. No latency benchmark has been recorded for this variant yet, so the latency column is left empty. Run build/bin/rfdetr-cli bench --model <gguf> --input <image> locally for timings on your own hardware.

Architecture

• Backbone: DINOv2-small
• Input resolution: 768×768
• Patch size: 12
• Decoder layers: 6
• Object queries: 300
• Task: instance segmentation (boxes + per-query masks)
• Mask resolution: 192×192 per query (image_size / 4)

Quantization notes

• F32 — the full-precision conversion, 144 MB, and the closest match to the PyTorch reference (see the table above for measured agreement).
• F16 — matmul-multiplicand weights only; LayerNorms, conv kernels, embeddings, biases, and layer-scale gammas stay F32. Accuracy tracks F32 closely on this model, and it takes ggml's F32×F16 matmul fast path.
• Q8_0 — best size/accuracy tradeoff under F16; ~3.0× smaller than F32 with effectively identical detections.
• Q4_K — smallest practical quant. Rows with ne[0] % 256 != 0 (the decoder's 128-dim MLP halves) silently fall back to Q8_0 per ggml's quantizer logic — net compression is still ~3.7× over F32. Use only when the size budget is tight; expect a measurable Recall@0.95 drop relative to F16/Q8_0 (see file table above).

Usage

# 1. Clone + build rfdetr.cpp
git clone https://github.com/mudler/rf-detr.cpp
cd rf-detr.cpp
cmake -B build -DRFDETR_BUILD_CLI=ON && cmake --build build -j

# 2. Download a quant (F16 recommended)
hf download mudler/rfdetr-cpp-seg-2xlarge rfdetr-seg-2xlarge-f16.gguf --local-dir models/

# 3. Run segmentation (writes per-detection PNG masks to /tmp/seg_masks/)
build/bin/rfdetr-cli detect \
    --model models/rfdetr-seg-2xlarge-f16.gguf \
    --input my_image.jpg \
    --threshold 0.5 --threads 8 \
    --masks /tmp/seg_masks \
    --output detections.json

Accuracy methodology

All accuracy metrics are computed against the upstream PyTorch reference (rfdetr 1.9.0) on 7 COCO val2017 images at threshold 0.5. Each detection match uses greedy Hungarian-style assignment by IoU (≥ 0.5 lenient, ≥ 0.95 strict) with class equality required.

Mask metrics are pixel-wise IoU between binary masks at the original image resolution (not the network's working resolution), after sigmoid + bicubic upsample of the per-query mask logits. Pixel agreement is the fraction of pixels where the C++ and PyTorch binary masks match.

See BENCHMARK.md and benchmarks/results/accuracy_sweep.json for the full sweep across the (variant × quant) cells.

License

Apache-2.0 — matches the upstream rfdetr license.