sebastientaylor commited on
Commit
7dec0ab
Β·
verified Β·
1 Parent(s): 9271eb4

Update model card for yolo11-seg

Browse files
Files changed (1) hide show
  1. README.md +103 -158
README.md CHANGED
@@ -9,14 +9,11 @@ tags:
9
  - onnx
10
  - int8
11
  - yolo
12
- - gstreamer
13
  - edgefirst
14
  - nxp
15
  - hailo
16
  - jetson
17
- - real-time
18
  - embedded
19
- - multiplatform
20
  model-index:
21
  - name: yolo11-seg
22
  results:
@@ -26,242 +23,190 @@ model-index:
26
  name: COCO val2017
27
  type: coco
28
  metrics:
 
 
 
29
  - name: "Mask mAP@0.5-0.95 (Nano ONNX FP32)"
30
  type: map
31
- value: 35.5
32
- - name: "Mask mAP@0.5-0.95 (Nano TFLite INT8)"
 
 
 
33
  type: map
34
- value: 34.4
 
 
 
 
 
 
35
  ---
36
 
37
- # YOLO11 Segmentation β€” EdgeFirst Edge AI
 
 
38
 
39
- **NXP i.MX 8M Plus** | **NXP i.MX 93** | **NXP i.MX 95** | **NXP Ara240** | **RPi5 + Hailo-8/8L** | **NVIDIA Jetson**
40
- YOLO11 Segmentation models optimized for edge AI deployment across multiple hardware platforms. All sizes from Nano to XLarge, in ONNX FP32 and TFLite INT8 formats, with platform-specific compiled models for NPU acceleration.
41
 
42
- Trained on [COCO 2017](https://test.edgefirst.studio/public/projects/2839/home) (80 classes). Part of the [EdgeFirst Model Zoo](https://huggingface.co/spaces/EdgeFirst/Models).
43
  > [!TIP]
44
- > **Training session**: [View on EdgeFirst Studio](https://test.edgefirst.studio/public/projects/2840/experiment/training/list?exp_id=4622) β€” dataset, training config, metrics, and exported artifacts.
45
 
46
  > [!NOTE]
47
- > Newer architecture with attention blocks.
48
-
49
- ---
50
-
51
- ## Size Comparison
52
-
53
- All models validated on COCO val2017 (5000 images, 80 classes).
54
-
55
- | Size | Params | GFLOPs | ONNX Det mAP@0.5-0.95 | INT8 Det mAP@0.5-0.95 | ONNX Mask mAP@0.5-0.95 | INT8 Mask mAP@0.5-0.95 |
56
- |------|--------|--------|------------------------|-----------------------|-------------------------|------------------------|
57
- | Nano | 2.6M | 6.5 | 28.4% | 27.1% | 35.5% | 34.4% |
58
- | Small | 9.4M | 21.5 | β€” | β€” | β€” | β€” |
59
- | Medium | 20.1M | 68.0 | β€” | β€” | β€” | β€” |
60
- | Large | 25.3M | 87.6 | β€” | β€” | β€” | β€” |
61
- | XLarge | 56.9M | 195.0 | β€” | β€” | β€” | β€” |
62
 
63
  ---
64
 
65
- ## On-Target Performance
66
 
67
- Full pipeline timing: pre-processing + inference + post-processing.
68
 
69
- | Size | Platform | Pre-proc (ms) | Inference (ms) | Post-proc (ms) | Total (ms) | FPS |
70
- |------|----------|---------------|----------------|-----------------|------------|-----|
71
- | β€” | β€” | β€” | β€” | β€” | β€” | β€” |
72
-
73
- *Measured with [EdgeFirst Perception](https://github.com/EdgeFirstAI) stack. Timing includes full GStreamer pipeline overhead.*
 
 
74
 
75
  ---
76
 
77
- ## Downloads
78
-
79
- <details open>
80
- <summary><strong>ONNX FP32</strong> β€” Any platform with ONNX Runtime.</summary>
81
-
82
- | Size | File | Status |
83
- |------|------|--------|
84
- | Nano | `yolo11n-seg-coco.onnx` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/onnx/yolo11n-seg-coco.onnx) |
85
- | Small | `yolo11s-seg-coco.onnx` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/onnx/yolo11s-seg-coco.onnx) |
86
- | Medium | `yolo11m-seg-coco.onnx` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/onnx/yolo11m-seg-coco.onnx) |
87
- | Large | `yolo11l-seg-coco.onnx` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/onnx/yolo11l-seg-coco.onnx) |
88
- | XLarge | `yolo11x-seg-coco.onnx` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/onnx/yolo11x-seg-coco.onnx) |
89
-
90
- </details>
91
-
92
- <details>
93
- <summary><strong>TFLite INT8</strong> β€” CPU or NPU via runtime delegate (i.MX 8M Plus VX Delegate).</summary>
94
-
95
- | Size | File | Status |
96
- |------|------|--------|
97
- | Nano | `yolo11n-seg-coco.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/tflite/yolo11n-seg-coco.tflite) |
98
- | Small | `yolo11s-seg-coco.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/tflite/yolo11s-seg-coco.tflite) |
99
- | Medium | `yolo11m-seg-coco.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/tflite/yolo11m-seg-coco.tflite) |
100
- | Large | `yolo11l-seg-coco.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/tflite/yolo11l-seg-coco.tflite) |
101
- | XLarge | `yolo11x-seg-coco.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/tflite/yolo11x-seg-coco.tflite) |
102
 
103
- </details>
104
 
105
- <details>
106
- <summary><strong>NXP i.MX 95 (eIQ Neutron)</strong> β€” eIQ Neutron NPU optimized.</summary>
107
-
108
- | Size | File | Status |
109
- |------|------|--------|
110
- | Nano | `yolo11n-seg-coco.imx95.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/imx95/yolo11n-seg-coco.imx95.tflite) |
111
- | Small | `yolo11s-seg-coco.imx95.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/imx95/yolo11s-seg-coco.imx95.tflite) |
112
- | Medium | `yolo11m-seg-coco.imx95.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/imx95/yolo11m-seg-coco.imx95.tflite) |
113
- | Large | `yolo11l-seg-coco.imx95.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/imx95/yolo11l-seg-coco.imx95.tflite) |
114
- | XLarge | `yolo11x-seg-coco.imx95.tflite` | [Download](https://huggingface.co/EdgeFirst/yolo11-seg/resolve/main/imx95/yolo11x-seg-coco.imx95.tflite) |
115
-
116
- </details>
117
 
 
 
 
 
 
 
 
118
 
119
 
120
  ---
121
 
122
- ## Deploy with EdgeFirst Perception
123
 
124
- Copy-paste [GStreamer](https://github.com/EdgeFirstAI/gstreamer) pipeline examples for each platform.
125
 
126
- ### NXP i.MX 8M Plus β€” Camera to Detection with Vivante NPU
 
 
127
 
128
- ```bash
129
- gst-launch-1.0 \
130
- v4l2src device=/dev/video0 ! video/x-raw,width=640,height=480 ! \
131
- edgefirstcameraadaptor ! \
132
- tensor_filter framework=tensorflow-lite \
133
- model=yolo11n-seg-coco.tflite \
134
- custom=Delegate:External,ExtDelegateLib:libvx_delegate.so ! \
135
- edgefirstsegdecoder ! edgefirstoverlay ! waylandsink
136
- ```
137
 
138
- ### RPi5 + Hailo-8L
139
 
140
- ```bash
141
- gst-launch-1.0 \
142
- v4l2src device=/dev/video0 ! video/x-raw,width=640,height=480 ! \
143
- hailonet hef-path=yolo11n-seg-coco.hailo8l.hef ! \
144
- hailofilter function-name=yolo11_nms ! \
145
- hailooverlay ! videoconvert ! autovideosink
146
- ```
147
-
148
- ### NVIDIA Jetson (TensorRT)
149
 
150
- ```bash
151
- gst-launch-1.0 \
152
- v4l2src device=/dev/video0 ! video/x-raw,width=640,height=480 ! \
153
- edgefirstcameraadaptor ! \
154
- nvinfer config-file-path=yolo11n-seg-coco-config.txt ! \
155
- edgefirstsegdecoder ! edgefirstoverlay ! nveglglessink
156
- ```
157
 
 
158
 
159
- *Full pipeline documentation: [EdgeFirst GStreamer Plugins](https://github.com/EdgeFirstAI/gstreamer)*
160
 
161
  ---
162
 
163
- ## Foundation (HAL) Python Integration
164
 
165
  ```python
166
  from edgefirst.hal import Model, TensorImage
167
 
168
- # Load model β€” metadata (labels, decoder config) is embedded in the file
169
- model = Model("yolo11n-seg-coco.tflite")
170
 
171
  # Run inference on an image
172
  image = TensorImage.from_file("image.jpg")
173
  results = model.predict(image)
174
 
175
- # Access detections
176
  for det in results.detections:
177
  print(f"{det.label}: {det.confidence:.2f} at {det.bbox}")
178
  ```
179
 
180
- *[EdgeFirst HAL](https://github.com/EdgeFirstAI/hal) β€” Hardware abstraction layer with accelerated inference delegates.*
181
 
182
  ---
183
 
184
- ## CameraAdaptor
185
-
186
- EdgeFirst [CameraAdaptor](https://github.com/EdgeFirstAI/cameraadaptor) enables training and inference directly on native sensor formats (GREY, YUYV, etc.) β€” skipping the ISP color conversion pipeline entirely. This reduces latency and power consumption on edge devices.
187
 
188
- CameraAdaptor variants are included alongside baseline RGB models:
189
 
190
- | Variant | Input Format | Use Case |
191
- |---------|-------------|----------|
192
- | `yolo11n-seg-coco.onnx` | RGB (3ch) | Standard camera input |
193
- | `yolo11n-seg-coco-grey.onnx` | GREY (1ch) | Monochrome / IR sensors |
194
- | `yolo11n-seg-coco-yuyv.onnx` | YUYV (2ch) | Raw sensor bypass |
195
 
196
- *Train CameraAdaptor models with [EdgeFirst Studio](https://edgefirst.studio) β€” the CameraAdaptor layer is automatically inserted during training.*
197
 
198
  ---
199
 
200
- ## Train Your Own with EdgeFirst Studio
201
 
202
- Train on your own dataset with [**EdgeFirst Studio**](https://edgefirst.studio):
203
 
204
- - **Free tier** includes YOLO training with automatic INT8 quantization and edge deployment
205
- - Upload datasets via [EdgeFirst Recorder](https://github.com/EdgeFirstAI/recorder) or COCO/YOLO format
206
- - AI-assisted annotation with auto-labeling
207
- - CameraAdaptor integration for native sensor format training
208
- - Deploy trained models to edge devices via [EdgeFirst Client](https://github.com/EdgeFirstAI/client)
 
 
 
209
 
210
  ---
211
 
212
- ## See Also
213
 
214
- Other models in the [EdgeFirst Model Zoo](https://huggingface.co/spaces/EdgeFirst/Models):
215
 
216
- | Model | Task | Best Nano Metric | Link |
217
- |-------|------|-------------------|------|
218
- | YOLOv5 Detection | Detection | 49.6% mAP@0.5 (ONNX) | [EdgeFirst/yolov5-det](https://huggingface.co/EdgeFirst/yolov5-det) |
219
- | YOLOv8 Detection | Detection | 50.2% mAP@0.5 (ONNX) | [EdgeFirst/yolov8-det](https://huggingface.co/EdgeFirst/yolov8-det) |
220
- | YOLOv8 Segmentation | Segmentation | 34.1% Mask mAP@0.5-0.95 (ONNX) | [EdgeFirst/yolov8-seg](https://huggingface.co/EdgeFirst/yolov8-seg) |
221
- | YOLO11 Detection | Detection | 53.4% mAP@0.5 (ONNX) | [EdgeFirst/yolo11-det](https://huggingface.co/EdgeFirst/yolo11-det) |
222
- | YOLO26 Detection | Detection | 54.9% mAP@0.5 (ONNX) | [EdgeFirst/yolo26-det](https://huggingface.co/EdgeFirst/yolo26-det) |
223
- | YOLO26 Segmentation | Segmentation | 37.0% Mask mAP@0.5-0.95 (ONNX) | [EdgeFirst/yolo26-seg](https://huggingface.co/EdgeFirst/yolo26-seg) |
224
 
225
  ---
226
 
227
- ## Technical Details
 
 
228
 
229
- ### Quantization Pipeline
230
 
231
- All TFLite INT8 models are produced by EdgeFirst's custom quantization pipeline ([details](https://github.com/EdgeFirstAI/studio-ultralytics)):
 
 
 
 
232
 
233
- 1. **ONNX Export** β€” Standard Ultralytics export with `simplify=True`
234
- 2. **TF-Wrapped ONNX** β€” Box coordinates normalized to [0,1] inside DFL decode via `tf_wrapper` (~1.2% better mAP than post-hoc normalization)
235
- 3. **Split Decoder** β€” Boxes, scores, and mask coefficients split into separate output tensors for independent INT8 quantization scales
236
- 4. **Smart Calibration** β€” 500 images selected via greedy coverage maximization from COCO val2017
237
- 5. **Full INT8** β€” `uint8` input (raw pixels), `int8` output (per-tensor scales), MLIR quantizer
238
 
239
- ### Split Decoder Output Format
240
 
241
- **Segmentation** (e.g., yolo11n-seg):
242
- - Boxes: `(1, 4, 8400)` β€” normalized [0,1] coordinates
243
- - Scores: `(1, 80, 8400)` β€” class probabilities
244
- - Mask coefficients: `(1, 32, 8400)` β€” per-anchor mask coefs
245
- - Protos: `(1, 160, 160, 32)` β€” prototype masks
246
 
247
- Each tensor has independent quantization scale and zero-point. EdgeFirst HAL handles dequantization and reassembly automatically.
248
 
249
- ### Metadata
250
 
251
- - **TFLite**: `edgefirst.json`, `labels.txt`, and `edgefirst.yaml` embedded via ZIP (no `tflite-support` dependency)
252
- - **ONNX**: `edgefirst.json` embedded via `model.metadata_props`
253
 
254
- No standalone metadata files β€” models are self-contained.
255
 
256
  ---
257
 
258
  ## Limitations
259
 
260
- - **COCO bias** β€” Models trained on COCO (80 classes) inherit its biases: Western-centric scenes, specific object distributions, limited weather/lighting diversity
261
- - **INT8 accuracy loss** β€” Full-integer quantization typically degrades mAP by 6-12% relative to FP32; actual loss depends on model architecture and dataset
262
- - **Thermal variation** β€” On-target performance varies with device temperature; sustained inference may throttle on passively-cooled devices
263
- - **Input resolution** β€” All models expect 640Γ—640 input; other resolutions require letterboxing or may reduce accuracy
264
- - **CameraAdaptor variants** β€” GREY/YUYV models trade color information for latency; accuracy may differ from RGB baseline depending on the task
265
 
266
  ---
267
 
@@ -269,7 +214,7 @@ No standalone metadata files β€” models are self-contained.
269
 
270
  ```bibtex
271
  @software{edgefirst_yolo11_seg,
272
- title = { {YOLO11 Segmentation β€” EdgeFirst Edge AI} },
273
  author = {Au-Zone Technologies},
274
  url = {https://huggingface.co/EdgeFirst/yolo11-seg},
275
  year = {2026},
 
9
  - onnx
10
  - int8
11
  - yolo
 
12
  - edgefirst
13
  - nxp
14
  - hailo
15
  - jetson
 
16
  - embedded
 
17
  model-index:
18
  - name: yolo11-seg
19
  results:
 
23
  name: COCO val2017
24
  type: coco
25
  metrics:
26
+ - name: "Box mAP@0.5 (Nano ONNX FP32)"
27
+ type: map_50
28
+ value: 52.0
29
  - name: "Mask mAP@0.5-0.95 (Nano ONNX FP32)"
30
  type: map
31
+ value: 28.75
32
+ - name: "Box mAP@0.5 (Small ONNX FP32)"
33
+ type: map_50
34
+ value: 60.6
35
+ - name: "Mask mAP@0.5-0.95 (Small ONNX FP32)"
36
  type: map
37
+ value: 33.33
38
+ - name: "Box mAP@0.5 (Medium ONNX FP32)"
39
+ type: map_50
40
+ value: 65.91
41
+ - name: "Mask mAP@0.5-0.95 (Medium ONNX FP32)"
42
+ type: map
43
+ value: 36.61
44
  ---
45
 
46
+ # YOLO11 Segmentation β€” EdgeFirst Model Zoo
47
+
48
+ YOLO11 Segmentation models trained on [COCO 2017](https://test.edgefirst.studio/public/projects/2839/home) (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.
49
 
50
+ Part of the [EdgeFirst Model Zoo](https://huggingface.co/spaces/EdgeFirst/Models).
 
51
 
 
52
  > [!TIP]
53
+ > **Training experiment**: [View on EdgeFirst Studio](https://test.edgefirst.studio/public/projects/2840/experiment/training/list?exp_id=4622) β€” dataset, training configuration, metrics, and exported artifacts.
54
 
55
  > [!NOTE]
56
+ > Architecture with C3k2 attention blocks.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
57
 
58
  ---
59
 
60
+ ## Reference accuracy β€” ONNX FP32
61
 
62
+ 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.
63
 
64
+ | Size | Params | GFLOPs | Box mAP@0.5 | Box mAP@0.5-0.95 | Mask mAP@0.5 | Mask mAP@0.5-0.95 | Source |
65
+ |------|--------|--------|-------------|------------------|--------------|-------------------|--------|
66
+ | Nano | 2.6M | 6.5 | 52.00% | 37.22% | 47.88% | 28.75% | [v-1d44](https://test.edgefirst.studio/public/validation/v-1d44/details?mode=metrics) |
67
+ | Small | 9.4M | 21.5 | 60.60% | 44.56% | 55.46% | 33.33% | [v-1d49](https://test.edgefirst.studio/public/validation/v-1d49/details?mode=metrics) |
68
+ | Medium | 20.1M | 68.0 | 65.91% | 49.59% | 60.54% | 36.61% | [v-1d4c](https://test.edgefirst.studio/public/validation/v-1d4c/details?mode=metrics) |
69
+ | Large | 25.3M | 87.6 | β€” | β€” | β€” | β€” | β€” |
70
+ | XLarge | 56.9M | 195.0 | β€” | β€” | β€” | β€” | β€” |
71
 
72
  ---
73
 
74
+ ## On-target validation results
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75
 
76
+ Each row is one EdgeFirst Studio validation session. Click the **Source** link to inspect the full session β€” model artifact, dataset version, parameters, per-stage Perfetto trace, and the host hardware description (hostname, kernel version, SoC, NPU, profiler version).
77
 
78
+ Cells rendered as `β€”` are sessions that did not meet the EdgeFirst publication threshold; the underlying session is still linked in the Source column for inspection.
 
 
 
 
 
 
 
 
 
 
 
79
 
80
+ | Size | Platform | Box mAP@0.5 | Mask mAP@0.5-0.95 | Ξ” mask vs FP32 (pp) | Inference (ms) | FPS (median) | Source |
81
+ |------|----------|-------------|-------------------|---------------------|----------------|--------------|--------|
82
+ | Nano | NXP i.MX 8M Plus + VeriSilicon NPU | 48.22% | 26.08% | -2.67 | 105.78 | 7.6 | [v-1d5f](https://test.edgefirst.studio/public/validation/v-1d5f/details?mode=metrics) |
83
+ | Nano | NXP i.MX 95 + eIQ Neutron NPU | β€” | β€” | β€” | β€” | β€” | [v-1d5e](https://test.edgefirst.studio/public/validation/v-1d5e/details?mode=metrics) |
84
+ | Nano | Raspberry Pi 5 + Hailo-8L NPU | 50.35% | 27.64% | -1.11 | 28.52 | 33.6 | [v-1d5c](https://test.edgefirst.studio/public/validation/v-1d5c/details?mode=metrics) |
85
+ | Small | Raspberry Pi 5 + Hailo-8L NPU | 58.66% | 32.39% | -0.94 | 40.49 | 24.0 | [v-1d5d](https://test.edgefirst.studio/public/validation/v-1d5d/details?mode=metrics) |
86
+ | Medium | Raspberry Pi 5 + Hailo-8L NPU | 63.95% | 35.52% | -1.09 | 83.81 | 11.8 | [v-1d60](https://test.edgefirst.studio/public/validation/v-1d60/details?mode=metrics) |
87
 
88
 
89
  ---
90
 
91
+ ## Validation pipeline
92
 
93
+ These results are produced by the EdgeFirst on-target validation pipeline:
94
 
95
+ 1. **EdgeFirst Profiler** runs on the target hardware, executes the full inference pipeline (preprocess β†’ inference β†’ postprocess), and emits per-image predictions in EdgeFirst Arrow/Parquet plus a Perfetto trace.
96
+ 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.
97
+ 3. **EdgeFirst HAL** ([open source](https://github.com/EdgeFirstAI/hal)) 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.
98
 
99
+ 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.
 
 
 
 
 
 
 
 
100
 
101
+ See [EdgeFirst Studio](https://edgefirst.studio) for the full validation pipeline.
102
 
103
+ ---
 
 
 
 
 
 
 
 
104
 
105
+ ## Downloads
 
 
 
 
 
 
106
 
107
+ 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 β€” no sidecar configuration required.
108
 
109
+ *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.*
110
 
111
  ---
112
 
113
+ ## Inference example (Python)
114
 
115
  ```python
116
  from edgefirst.hal import Model, TensorImage
117
 
118
+ # Load the model β€” embedded edgefirst.json carries labels and decoder config
119
+ model = Model("yolo11n-seg-int8.tflite")
120
 
121
  # Run inference on an image
122
  image = TensorImage.from_file("image.jpg")
123
  results = model.predict(image)
124
 
125
+ # Iterate detections
126
  for det in results.detections:
127
  print(f"{det.label}: {det.confidence:.2f} at {det.bbox}")
128
  ```
129
 
130
+ [EdgeFirst HAL](https://github.com/EdgeFirstAI/hal) β€” Hardware abstraction layer with accelerated inference delegates.
131
 
132
  ---
133
 
134
+ ## Traceability
 
 
135
 
136
+ 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:
137
 
138
+ ```
139
+ https://test.edgefirst.studio/public/validation/v-XXXX/details?mode=metrics
140
+ ```
 
 
141
 
142
+ 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.
143
 
144
  ---
145
 
146
+ ## See also
147
 
148
+ Other model families in the [EdgeFirst Model Zoo](https://huggingface.co/spaces/EdgeFirst/Models):
149
 
150
+ | Model | Task | Link |
151
+ |-------|------|------|
152
+ | YOLOv5 Detection | Detection | [EdgeFirst/yolov5-det](https://huggingface.co/EdgeFirst/yolov5-det) |
153
+ | YOLOv8 Detection | Detection | [EdgeFirst/yolov8-det](https://huggingface.co/EdgeFirst/yolov8-det) |
154
+ | YOLOv8 Segmentation | Segmentation | [EdgeFirst/yolov8-seg](https://huggingface.co/EdgeFirst/yolov8-seg) |
155
+ | YOLO11 Detection | Detection | [EdgeFirst/yolo11-det](https://huggingface.co/EdgeFirst/yolo11-det) |
156
+ | YOLO26 Detection | Detection | [EdgeFirst/yolo26-det](https://huggingface.co/EdgeFirst/yolo26-det) |
157
+ | YOLO26 Segmentation | Segmentation | [EdgeFirst/yolo26-seg](https://huggingface.co/EdgeFirst/yolo26-seg) |
158
 
159
  ---
160
 
161
+ ## Train your own with EdgeFirst Studio
162
 
163
+ Train on your own dataset with [**EdgeFirst Studio**](https://edgefirst.studio):
164
 
165
+ - Free tier includes YOLO training with automatic INT8 quantization and edge deployment.
166
+ - Upload datasets via [EdgeFirst Recorder](https://github.com/EdgeFirstAI/recorder) or COCO/YOLO format.
167
+ - AI-assisted annotation with auto-labeling.
168
+ - CameraAdaptor integration for native sensor format training.
169
+ - Deploy trained models to edge devices via [EdgeFirst Client](https://github.com/EdgeFirstAI/client).
 
 
 
170
 
171
  ---
172
 
173
+ ## Technical notes
174
+
175
+ ### Quantization pipeline
176
 
177
+ All TFLite INT8 models are produced by EdgeFirst's quantization pipeline ([details](https://github.com/EdgeFirstAI/studio-ultralytics)):
178
 
179
+ 1. **ONNX export** β€” standard Ultralytics export with `simplify=True`
180
+ 2. **TF-wrapped ONNX** β€” box coordinates normalized to `[0, 1]` inside DFL decode
181
+ 3. **Split decoder** β€” boxes, scores, and mask coefficients split into separate output tensors so each receives an independent INT8 quantization scale
182
+ 4. **Smart calibration** β€” calibration samples selected via greedy coverage maximization; the artifact is content-addressed by parameter hash and cached in Studio for deterministic reuse
183
+ 5. **Full integer INT8** β€” `uint8` input, `int8` output, MLIR quantizer
184
 
185
+ ### Split decoder output format
 
 
 
 
186
 
187
+ **Segmentation** (e.g. yolo11n-seg):
188
 
189
+ - `boxes` β€” `(1, 4, 8400)` normalized `[0, 1]` coordinates
190
+ - `scores` β€” `(1, 80, 8400)` per-class probabilities
191
+ - `mask_coefs` β€” `(1, 32, 8400)` per-anchor mask coefficients
192
+ - `protos` β€” `(1, 160, 160, 32)` prototype masks
 
193
 
194
+ 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.
195
 
196
+ ### Embedded metadata
197
 
198
+ - **TFLite**: `edgefirst.json` and `labels.txt` embedded in the ZIP-format model file
199
+ - **ONNX**: `edgefirst.json` embedded in `model.metadata_props`
200
 
201
+ No sidecar files required; the model artifact is self-contained.
202
 
203
  ---
204
 
205
  ## Limitations
206
 
207
+ - **COCO bias** β€” models trained on COCO (80 classes) inherit the dataset's biases (Western-centric scenes, particular object distributions, limited weather/lighting diversity).
208
+ - **INT8 quantization loss** β€” full-integer quantization introduces accuracy loss relative to FP32; the magnitude per platform is shown in the *Ξ” vs FP32* column above.
209
+ - **Input resolution** β€” all models expect 640Γ—640 input; other resolutions require letterboxing.
 
 
210
 
211
  ---
212
 
 
214
 
215
  ```bibtex
216
  @software{edgefirst_yolo11_seg,
217
+ title = { {YOLO11 Segmentation β€” EdgeFirst Model Zoo} },
218
  author = {Au-Zone Technologies},
219
  url = {https://huggingface.co/EdgeFirst/yolo11-seg},
220
  year = {2026},