YOLO26-seg: Optimized for AMD ROCm

YOLO26 is a real-time instance segmentation model that detects objects and produces pixel-level masks across 80 COCO categories in a single forward pass. This repository packages inference for instance segmentation using ONNX Runtime (CPU, NPU) and PyTorch (GPU), exported and validated for AMD ROCm so it runs efficiently on AMD GPUs, CPUs, and NPUs.

This is based on the implementation of YOLO26 found here. This repository contains configurations and scripts optimized for AMD® ROCm™ platforms. You can use the yolo26_segment AMD scripts to reproduce results or export with custom configurations. More details on model performance can be found here.

Task Overview

Task: Instance segmentation

Dataset: COCO val2017 (5,000 images, 80 categories)

Output metrics: Mask mAP@0.5:0.95, Mask mAP@0.5, Mask mAP@0.75, Bbox mAP, Mask Precision, Mask Recall, Mask F1, per-size Mask mAP (small/medium/large)

Model variants: Default is n. Override with MODEL_SIZE=n/s/m/l/x.

GPU note: GPU quality uses PyTorch .pt models directly — MIGraphX does not support YOLO segment ONNX ops, so GPU runs FP32 only via PyTorch rather than MIGraphX.

AMD ROCm Optimization

This model export has been adapted and validated for AMD Instinct™ / Radeon™ GPUs running ROCm, as well as AMD CPUs and AMD Ryzen AI NPUs. Key points:

Validated backends: ONNX Runtime on CPU (FP32) and NPU (VitisAI — auto-quantized internally, BF16), and PyTorch/ROCm on GPU (FP32 only).
No code changes required versus the upstream Ultralytics YOLO26-seg implementation — only environment/runtime configuration differs.
First NPU run takes 5–10 minutes for model compilation; subsequent runs use the cached compiled model.

Runtime	Precision	Backend	Hardware	Notes
ONNX Runtime	FP32	CPU Execution Provider	AMD CPU	—
PyTorch	FP32	ROCm (HIP)	AMD Instinct™ / Radeon™ GPU	MIGraphX does not support segment ONNX ops
ONNX Runtime	BF16	VitisAI Execution Provider	AMD Ryzen AI NPU	Auto-quantized internally; first run compiles in 5–10 minutes

Getting Started

For setup instructions, evaluation scripts, and custom configuration options, see the yolo26_segment on GitHub.

Model Details

Model Type: Instance segmentation (single-pass CNN detector with mask-proto head)

Base Model: YOLO26-seg (Ultralytics), size n default

Model Stats:

Input: (1, 3, 640, 640) float32
Output (detections): (1, 300, 38) float32
Output (mask protos): (1, 32, 160, 160) float32
Precision tested: FP32 (CPU, GPU); BF16 (NPU)

Accuracy Pipeline

COCO instance segmentation quality evaluation is fully implemented — make eval-fulldataset-<device> runs inference across all 5,000 COCO val2017 images and computes official COCO metrics via pycocotools.

Higher mask mAP means the model's predicted segmentation masks agree more closely with ground-truth masks — 1.0 would be perfect pixel-level segmentation, 0.0 means no correct predictions. In practice, values above ~0.35 for mask mAP@0.5:0.95 are considered strong for COCO-scale instance segmentation.

Metrics Explained

Metric	Description
Mask mAP@0.5:0.95	Primary COCO mask metric — mean AP averaged across IoU thresholds 0.5–0.95. The strictest, most holistic segmentation accuracy number; higher means masks are both correctly classified and tightly overlap ground-truth at a range of overlap thresholds.
Mask mAP@0.5	Mask AP at a single, looser IoU threshold of 0.5 — a prediction only needs 50% pixel overlap with the ground-truth mask to count as correct, so this is typically higher than mAP@0.5:0.95 and reflects "did it find the object" more than "how precise is the mask boundary."
Mask mAP@0.75	Mask AP at a stricter IoU threshold of 0.75 — the predicted mask must overlap ground truth by 75%, rewarding precise boundary delineation, not just correct detection.
Bbox mAP	Bounding-box mAP@0.5:0.95 — measures detection quality (box localization) independently of mask quality, useful for isolating whether errors come from detection or segmentation.
Mask Precision	Of all mask predictions the model produced, what fraction were correct — high precision means few false positive masks.
Mask Recall	Of all ground-truth objects, what fraction did the model find within its top 100 detections per image — high recall means few missed objects.
Mask F1	Harmonic mean of mask precision and recall — balances the tradeoff between finding all objects and avoiding false detections.
Mask mAP-small/medium/large	Mask mAP@0.5:0.95 broken down by object size — small objects are typically the hardest to segment, exposing size-specific weaknesses a single aggregate score would hide.

Accuracy Results

Full Dataset Evaluation (COCO val2017) — MODEL_SIZE=n:

Device	Precision	Mask mAP@0.5:0.95	Mask mAP@0.5	Mask mAP@0.75	Bbox mAP	Mask Precision	Mask Recall
CPU	FP32	0.3340	0.5258	0.3528	0.3962	0.3031	0.4244
GPU	FP32	0.2757	0.4774	0.2777	0.3954	0.2534	0.3617
NPU	BF16	0.0008	0.0022	0.0005	0.0041	0.0007	0.0020

Note: GPU quality uses PyTorch .pt models directly (MIGraphX does not support segment ONNX ops). CPU mask mAP is higher than GPU because the ONNX post-NMS postprocessing applies tighter bbox-cropped masks. NPU quality is severely degraded by VitisAI's internal BF16 quantization — the model produces very few detections (bbox mAP drops from ~0.40 to 0.004), making NPU unsuitable for accuracy-critical workloads with this model.

Dig Deeper

Want to explore the full evaluation scripts, config options, and other AMD-optimized model examples?

📂 View the full project on GitHub

The GitHub repository includes:

Setup and prerequisites for ROCm environments
Full COCO val2017 evaluation pipeline via pycocotools
Per-operator latency profiling scripts (including NPU AI Analyzer integration)
Benchmarking and reproduction instructions across CPU, GPU, and NPU

Downloads last month: -