Hugging Face's logo

STMicroelectronics
/
osnet

Model card Files
xet
Community
FBAGSTM commited on
Commit
7e31643
·
verified ·
1 Parent(s): 7a4d95a

Release AI-ModelZoo-4.0.0

Browse files
Files changed (1) hide show
  1. README.md +121 -5
README.md CHANGED
@@ -1,5 +1,121 @@
1
- ---
2
- license: mit
3
- license_name: sla0044
4
- license_link: LICENSE
5
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ license_name: sla0044
4
+ license_link: LICENSE
5
+ ---
6
+ # OSNet
7
+
8
+ ## **Use case** : `Re-Identification`
9
+
10
+ # Model description
11
+
12
+
13
+ OSNet is a lightweight convolutional neural network architecture designed specifically for person re-identification tasks. It introduces omni-scale feature learning, enabling the network to capture multi-scale information efficiently within a single residual block.
14
+
15
+ Key features of OSNet:
16
+ - Omni-scale feature learning for robust representation.
17
+ - Lightweight design with fewer parameters compared to traditional re-identification models.
18
+ - Suitable for deployment on resource-constrained devices.
19
+
20
+ For more details, see the OSNet paper: https://arxiv.org/abs/1905.00953
21
+
22
+ The model is quantized using ONNX quantization tools.
23
+
24
+ ## Network information
25
+
26
+
27
+ | Network Information | Value |
28
+ |-------------------------|-----------------|
29
+ | Framework | TensorFlow Lite |
30
+ | MParams alpha=0.25 | 0.197 M |
31
+ | Quantization | int8 |
32
+ | Provenance | https://kaiyangzhou.github.io/deep-person-reid/index.html |
33
+ | Paper | https://arxiv.org/abs/1905.0095 |
34
+
35
+ The models are quantized using TF Lite post-training quantization tools.
36
+
37
+
38
+ ## Network inputs / outputs
39
+
40
+
41
+ For an image resolution of NxM and P classes
42
+
43
+ | Input Shape | Description |
44
+ | ----- | ----------- |
45
+ | (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
46
+
47
+ | Output Shape | Description |
48
+ | ----- | ----------- |
49
+ | (1, P) | Per-class confidence for P classes in FLOAT32|
50
+
51
+
52
+ ## Recommended platforms
53
+
54
+
55
+ | Platform | Supported | Recommended |
56
+ |----------|-----------|-----------|
57
+ | STM32L0 |[]|[]|
58
+ | STM32L4 |[x]|[]|
59
+ | STM32U5 |[x]|[]|
60
+ | STM32H7 |[x]|[x]|
61
+ | STM32MP1 |[x]|[x]|
62
+ | STM32MP2 |[x]|[x]|
63
+ | STM32N6 |[x]|[x]|
64
+
65
+ # Performances
66
+
67
+ ## Metricss
68
+
69
+ - Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
70
+ - `tfs` stands for "training from scratch", meaning that the model weights were randomly initialized before training.
71
+ - `tl` stands for "transfer learning", meaning that the model backbone weights were initialized from a pre-trained model, then only the last layer was unfrozen during the training.
72
+ - `fft` stands for "full fine-tuning", meaning that the full model weights were initialized from a transfer learning pre-trained model, and all the layers were unfrozen during the training.
73
+
74
+
75
+ ### Reference **NPU** memory footprint on DeepSportradar dataset (see Accuracy for details on dataset)
76
+ |Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
77
+ |----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------|-------------------------|
78
+ | [OSNet 0.25 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a025_256_128_tfs/osnet_a025_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6 | 480 | 0 | 404.94 | 3.0.0 |
79
+ | [OSNet 1.0 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a100_256_128_tfs/osnet_a100_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6 | 1440 | 0 | 2375.33 | 3.0.0 |
80
+
81
+
82
+ ### Reference **NPU** inference time on DeepSportradar dataset (see Accuracy for details on dataset)
83
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
84
+ |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------| -----------------------|
85
+ | [OSNet 0.25 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a025_256_128_tfs/osnet_a025_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6570-DK | NPU/MCU | 3.53 | 283.3 | 3.0.0 |
86
+ | [OSNet 1.0 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a100_256_128_tfs/osnet_a100_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32N6570-DK | NPU/MCU | 13.44 | 74.4 | 3.0.0 |
87
+
88
+
89
+ ### Reference **MCU** memory footprint based on DeepSportradar dataset (see Accuracy for details on dataset)
90
+ |Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
91
+ |----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------|-------------------------|
92
+ | [OSNet 0.25 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a025_256_128_tfs/osnet_a025_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H7 | 331.45 | 0 | 139.52 | 3.0.0 |
93
+ | [OSNet 1.0 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a100_256_128_tfs/osnet_a100_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H7 | 396.01 | 1024.0 | 1892.75 | 3.0.0 |
94
+
95
+ ### Reference **MCU** inference time on DeepSportradar dataset (see Accuracy for details on dataset)
96
+ | Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
97
+ |--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------| -----------------------|
98
+ | [OSNet 0.25 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a025_256_128_tfs/osnet_a025_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H747I-DISCO | 1 CPU | 495.13 | 2.02 | 3.0.0 |
99
+ | [OSNet 1.0 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a100_256_128_tfs/osnet_a100_256_128_tfs_int8.tflite) | DeepSportradar | Int8 | 256x128x3 | STM32H747I-DISCO | 1 CPU | 3894.82 | 0.26 | 3.0.0 |
100
+
101
+
102
+ ### Performance with DeepSportradar ReID dataset
103
+
104
+
105
+ Dataset details: [link](https://github.com/DeepSportradar/player-reidentification-challenge) , License [Apache-2.0](https://github.com/DeepSportradar/player-reidentification-challenge?tab=Apache-2.0-1-ov-file#readme) , Number of identities: 486 (train: 436, test: 50), Number of images: 9529 (train: 8569, test_query: 50, test_gallery: 910)
106
+
107
+ | Model | Format | Resolution | mAP | rank-1 accuracy |rank-5 accuracy |rank-10 accuracy |
108
+ |-------|--------|------------|----------------|-----------------|----------------|-----------------|
109
+ | [OSNet 0.25 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a025_256_128_tfs/osnet_a025_256_128_tfs_int8.tflite) | Int8 | 256x128 | 70.27 % | 92.0 % | 96.0 % | 96.0 % |
110
+ | [OSNet 1.0 tfs](https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/main/re_identification/osnet/ST_pretrainedmodel_public_dataset/DeepSportradar/osnet_a100_256_128_tfs/osnet_a100_256_128_tfs_int8.tflite) | Int8 | 256x128 | 73.84 % | 90.0 % | 98.0 % | 98.0 % |
111
+
112
+
113
+ ## Retraining and Integration in a simple example:
114
+
115
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
116
+
117
+
118
+ # References
119
+
120
+ <a id="1">[1]</a>
121
+ The DeepSportradar Player Re-Identification Challenge (2023) [Online]. Available: https://github.com/DeepSportradar/player-reidentification-challenge.