adithya-balaji/rfdetr-cpp-seg-medium

RF-DETR Seg-Medium — GGUF for rfdetr.cpp

GGUF-format weights of Roboflow RF-DETR Seg-Medium (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 — same accuracy as F32, 1.85× smaller, and typically the fastest on modern CPUs thanks to 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-medium-f32.gguf	F32	133.9	0.9841	0.9841	0.9978	0.9999	199.5
rfdetr-seg-medium-f16.gguf ← recommended	F16	71.6	0.9841	0.9841	0.9970	0.9999	208.5
rfdetr-seg-medium-q8_0.gguf	Q8_0	42.5	0.9841	0.9841	0.9935	0.9999	218.5
rfdetr-seg-medium-q4_K.gguf	Q4_K	33.6	0.9137	0.4787	0.9687	0.9993	244.2

All accuracy numbers above are computed against the upstream PyTorch reference (rfdetr 1.9.0) on 7 images (000000000139.jpg, 000000000632.jpg, 000000039769.jpg, 000000087038.jpg, 000000252219.jpg, 000000397133.jpg, bus.jpg) at threshold 0.5. Latency is measured separately with rfdetr-cli bench (8 iters + 3 warmup) at T=8 threads on a single Intel Core i7-12800HX, on tests/fixtures/ci/test_image.jpg.

Architecture

• Backbone: DINOv2-small
• Input resolution: 432×432
• Patch size: 12
• Decoder layers: 5
• Object queries: 200
• Task: instance segmentation (boxes + per-query masks)
• Mask resolution: 108×108 per query (image_size / 4)

Quantization notes

• F32 — full-precision reference, ~120 MB. Bit-exact PyTorch parity.
• F16 — matmul-multiplicand weights only; LayerNorms, conv kernels, embeddings, biases, and layer-scale gammas stay F32. Lossless on this model and consistently the fastest variant on CPU.
• Q8_0 — best size/accuracy tradeoff under F16; ~3× 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, 60 tensors) silently fall back to Q8_0 per ggml's quantizer logic — net compression is still ~3.8× 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).

Compatibility

These GGUFs stamp rfdetr.preprocess.resize_mode = "bilinear_no_antialias", matching RF-DETR 1.9's antialias-free float bilinear resize (align_corners=false, half-pixel coordinates, no intermediate uint8 rounding). rf-detr.cpp treats this key as optional: GGUFs that predate this metadata (no resize_mode key) keep using the legacy stb-based resize path, so older files continue to produce their original outputs unchanged. An unrecognized resize_mode value is rejected rather than guessed.

Keypoint-preview inference is not supported. rf-detr.cpp does not implement the keypoint output head; this repository only serves box detection + instance segmentation masks outputs.

Usage

# 1. Clone + build rfdetr.cpp
git clone https://github.com/adithyab94/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 adithya-balaji/rfdetr-cpp-seg-medium rfdetr-seg-medium-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-medium-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 images (000000000139.jpg, 000000000632.jpg, 000000039769.jpg, 000000087038.jpg, 000000252219.jpg, 000000397133.jpg, bus.jpg) 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.

Provenance

• Source project: Roboflow RF-DETR
• Upstream package: rfdetr==1.9.0
• Converted with rfdetr.cpp at commit fbef9387bed3
• Checkpoint: official pretrained rfdetr-seg-medium weights (downloaded by the rfdetr package on first use)

Checksums (SHA-256)

Also available as SHA256SUMS in this repo.

0b5594a96cfb3aebba698b536bf84205101034bad60b166d1cc3b7606adc3be3  rfdetr-seg-medium-f32.gguf
dd7c8da7cf0a2e64a1002f5ff66d7fede45b00e612457f249bcd9d4a0c122566  rfdetr-seg-medium-f16.gguf
1980f61f4ba57b874007a3ebc0685caeacb92171c68c70d31b604d973104a3d4  rfdetr-seg-medium-q8_0.gguf
461f2dec286926f5eda33d831f11689516039918e12f23891b7789ce36049841  rfdetr-seg-medium-q4_K.gguf

License

Apache-2.0 — matches the upstream rfdetr license.