EdgeFirst/yolov5-det

YOLOv5 Detection — EdgeFirst Model Zoo

YOLOv5 Detection models trained on COCO 2017 (80 classes) and validated on real edge hardware through the EdgeFirst Profiler + Validator pipeline. Each row in the tables below cites the EdgeFirst Studio validation session (v-XXXX) that produced the measurement.

Part of the EdgeFirst Model Zoo.

Training experiment: View on EdgeFirst Studio — dataset, training configuration, metrics, and exported artifacts.

Legacy architecture, wide deployment base.

---

Reference accuracy — ONNX FP32

Accuracy ceiling for each size, measured against COCO val2017 (5,000 images) with pycocotools. Quantized and compiled artifacts (TFLite INT8, HEF, etc.) are graded against this reference per the EdgeFirst publication rule.

Size	Params	GFLOPs	mAP@0.5	mAP@0.5-0.95	mAP@0.75	Source
Nano	1.9M	4.5	47.77%	32.93%	35.68%	v-1d56
Small	7.2M	16.5	57.33%	41.26%	44.93%	v-1db0
Medium	21.2M	49.0	63.31%	47.00%	51.38%	v-1db2
Large	46.5M	109.1	—	—	—	—
XLarge	86.7M	205.7	—	—	—	—

---

On-target validation results

Each row is one EdgeFirst Studio validation session. Click the Source link to inspect the full session — model artifact, dataset version, parameters, per-stage Perfetto trace, and the host hardware description (hostname, kernel version, SoC, NPU, profiler version).

Two row conventions in the table below:

• Rows without a number under the metric columns are validation sessions that are currently work in progress — typically a larger size that has not yet been profiled on a given NPU, or a session that has not yet been linked to an ONNX FP32 reference. The Studio Source link tracks the current status.
• Rows whose Δ vs FP32 cell carries a ⚠ are published with a caveat: the quantization or runtime regression on that platform exceeds the 10 percentage-point publication threshold we grade every measurement against. The numbers are still measured on real hardware and reproducible from the linked Studio session — they are shown for transparency rather than suppressed. We are actively investigating these regressions in collaboration with the toolchain vendor and working to close the gap; the note immediately under the table summarises the current state.

Size	Platform	mAP@0.5	Δ vs FP32 (pp)	mAP@0.5-0.95	Inference (ms)	FPS (median)	Source
Nano	ONNX FP32 (CPU)	47.77%	+0.00	32.93%	5.45	161.6	v-1d2d
Nano	ONNX FP32	47.75%	-0.02	32.92%	2.89	342.8	v-1dae
Nano	ONNX FP32	47.70%	-0.07	32.89%	5.00	198.1	v-1dad
Nano	NXP i.MX 8M Plus + VeriSilicon NPU	46.56%	-1.21	31.92%	58.38	12.6	v-1d2b
Nano	NXP i.MX 95 + eIQ Neutron NPU	46.02%	-1.75	31.19%	11.95	58.7	v-1d2c
Nano	Raspberry Pi 5 + Hailo-8L NPU	45.49%	-2.28	31.17%	15.68	61.3	v-1d26
Nano	(no-footer)	0.00%	-47.77 ⚠	0.00%	15.04	59.8	v-1dba
Small	ONNX FP32 (CPU)	57.32%	-0.01	41.27%	6.02	155.1	v-1d2e
Small	ONNX FP32	57.28%	-0.05	41.15%	7.67	128.8	v-1db1
Small	NXP i.MX 8M Plus + VeriSilicon NPU	56.34%	-0.99	40.20%	110.36	7.6	v-1d38
Small	NXP i.MX 95 + eIQ Neutron NPU	56.19%	-1.14	40.07%	31.91	27.1	v-1d32
Small	Raspberry Pi 5 + Hailo-8L NPU	55.95%	-1.38	40.06%	30.51	32.0	v-1d27
Small	(no-footer)	0.00%	-57.33 ⚠	0.00%	16.08	58.9	v-1dbb
Medium	ONNX FP32 (CPU)	63.29%	-0.02	47.00%	8.52	112.8	v-1d2f
Medium	ONNX FP32	63.29%	-0.02	47.00%	10.03	99.5	v-1db3
Medium	NXP i.MX 95 + eIQ Neutron NPU	62.74%	-0.57	46.52%	77.14	11.9	v-1d39
Medium	Raspberry Pi 5 + Hailo-8L NPU	62.50%	-0.81	46.24%	61.08	16.2	v-1d28

⚠ Investigations in progress. The rows marked ⚠ above show a drop of more than 10 percentage points against the same training session's ONNX FP32 reference. We publish them as-is so the current state of every (model, platform) pair is visible; we are working with the toolchain vendor to close each regression, and the numbers will improve as fixes land. The next snapshot of this card will reflect any recovered accuracy.

---

Validation pipeline

These results are produced by the EdgeFirst on-target validation pipeline:

1. EdgeFirst Profiler runs on the target hardware, executes the full inference pipeline (preprocess → inference → postprocess), and emits per-image predictions in EdgeFirst Arrow/Parquet plus a Perfetto trace.
2. EdgeFirst Validator consumes the predictions and trace, computes pycocotools accuracy metrics and per-stage timing summaries, and publishes the results to the Studio validation session.
3. EdgeFirst HAL (open source) provides the hardware-accelerated preprocessing and post-decoding primitives used at both validation and deployment time, so the timings measured here reflect the same accelerated paths a production runtime would take.

Inference latency is reported as the on-accelerator inference time. FPS is the measured end-to-end pipelined throughput from the Perfetto trace, which generally exceeds 1000 / (preprocess + inference + postprocess) because the runtime overlaps stages across frames.

See EdgeFirst Studio for the full validation pipeline.

---

Downloads

Artifacts are organized by deployment target. Each model file embeds the EdgeFirst edgefirst.json metadata (training session, dataset version, calibration artifact, converter chain) so a single file is sufficient for deployment — no sidecar configuration required.

Per-artifact download links are populated from the Studio artifact registry. To see the live download table, regenerate this card with --studio against an authenticated Studio session.

---

Inference example (Python)

from edgefirst.hal import Model, TensorImage

# Load the model — embedded edgefirst.json carries labels and decoder config
model = Model("yolov5n-det-int8.tflite")

# Run inference on an image
image = TensorImage.from_file("image.jpg")
results = model.predict(image)

# Iterate detections
for det in results.detections:
    print(f"{det.label}: {det.confidence:.2f} at {det.bbox}")

EdgeFirst HAL — Hardware abstraction layer with accelerated inference delegates.

---

Traceability

Every measurement in the tables above is reachable through the EdgeFirst Studio validation framework. The v-XXXX Source link on each row resolves to a public Studio URL of the form:

https://test.edgefirst.studio/public/validation/v-XXXX/details

From there, the full provenance chain is one click deeper: training session ID, dataset version, calibration artifact, converter chain (e.g. TFLite quantizer + Neutron compile), validation parameters, and the host hardware description (hostname, kernel version, SoC, NPU, profiler version). The same model file you download from this repository embeds the same chain in its edgefirst.json metadata.

---

See also

Other model families in the EdgeFirst Model Zoo:

Model	Task	Link
YOLOv8 Detection	Detection	EdgeFirst/yolov8-det
YOLOv8 Segmentation	Segmentation	EdgeFirst/yolov8-seg
YOLO11 Detection	Detection	EdgeFirst/yolo11-det
YOLO11 Segmentation	Segmentation	EdgeFirst/yolo11-seg
YOLO26 Detection	Detection	EdgeFirst/yolo26-det
YOLO26 Segmentation	Segmentation	EdgeFirst/yolo26-seg

---

Train your own with EdgeFirst Studio

Train on your own dataset with EdgeFirst Studio:

• Free tier includes YOLO training with automatic INT8 quantization and edge deployment.
• Upload datasets via EdgeFirst Recorder or COCO/YOLO format.
• AI-assisted annotation with auto-labeling.
• CameraAdaptor integration for native sensor format training.
• Deploy trained models to edge devices via EdgeFirst Client.

---

Technical notes

Quantization pipeline

All TFLite INT8 models are produced by EdgeFirst's quantization pipeline (details):

1. ONNX export — standard Ultralytics export with simplify=True
2. TF-wrapped ONNX — box coordinates normalized to [0, 1] inside DFL decode
3. Split decoder — boxes, scores, and mask coefficients split into separate output tensors so each receives an independent INT8 quantization scale
4. Smart calibration — calibration samples selected via greedy coverage maximization; the artifact is content-addressed by parameter hash and cached in Studio for deterministic reuse
5. Full integer INT8 — uint8 input, int8 output, MLIR quantizer

Split decoder output format

Detection (e.g. yolov5n):

• boxes — (1, 4, 8400) normalized [0, 1] coordinates
• scores — (1, 80, 8400) per-class probabilities

Each tensor has its own quantization scale and zero point. The EdgeFirst HAL handles dequantization and reassembly automatically; no application code change is required across NPU targets.

Embedded metadata

• TFLite: edgefirst.json and labels.txt embedded in the ZIP-format model file
• ONNX: edgefirst.json embedded in model.metadata_props

No sidecar files required; the model artifact is self-contained.

---

Limitations

• COCO bias — models trained on COCO (80 classes) inherit the dataset's biases (Western-centric scenes, particular object distributions, limited weather/lighting diversity).
• INT8 quantization loss — full-integer quantization introduces accuracy loss relative to FP32; the magnitude per platform is shown in the Δ vs FP32 column above.
• Input resolution — all models expect 640×640 input; other resolutions require letterboxing.

---

Citation

@software{edgefirst_yolov5_det,
  title = { {YOLOv5 Detection — EdgeFirst Model Zoo} },
  author = {Au-Zone Technologies},
  url = {https://huggingface.co/EdgeFirst/yolov5-det},
  year = {2026},
  license = {Apache-2.0},
}

---

_{EdgeFirst Studio · GitHub · Docs · Au-Zone Technologies
Apache 2.0 · © Au-Zone Technologies Inc.}