File size: 8,319 Bytes
7e31643 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 |
---
license: mit
license_name: sla0044
license_link: LICENSE
---
# OSNet
## **Use case** : `Re-Identification`
# Model description
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.
Key features of OSNet:
- Omni-scale feature learning for robust representation.
- Lightweight design with fewer parameters compared to traditional re-identification models.
- Suitable for deployment on resource-constrained devices.
For more details, see the OSNet paper: https://arxiv.org/abs/1905.00953
The model is quantized using ONNX quantization tools.
## Network information
| Network Information | Value |
|-------------------------|-----------------|
| Framework | TensorFlow Lite |
| MParams alpha=0.25 | 0.197 M |
| Quantization | int8 |
| Provenance | https://kaiyangzhou.github.io/deep-person-reid/index.html |
| Paper | https://arxiv.org/abs/1905.0095 |
The models are quantized using TF Lite post-training quantization tools.
## Network inputs / outputs
For an image resolution of NxM and P classes
| Input Shape | Description |
| ----- | ----------- |
| (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
| Output Shape | Description |
| ----- | ----------- |
| (1, P) | Per-class confidence for P classes in FLOAT32|
## Recommended platforms
| Platform | Supported | Recommended |
|----------|-----------|-----------|
| STM32L0 |[]|[]|
| STM32L4 |[x]|[]|
| STM32U5 |[x]|[]|
| STM32H7 |[x]|[x]|
| STM32MP1 |[x]|[x]|
| STM32MP2 |[x]|[x]|
| STM32N6 |[x]|[x]|
# Performances
## Metricss
- Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
- `tfs` stands for "training from scratch", meaning that the model weights were randomly initialized before training.
- `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.
- `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.
### Reference **NPU** memory footprint on DeepSportradar dataset (see Accuracy for details on dataset)
|Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
|----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------|-------------------------|
| [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 |
| [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 |
### Reference **NPU** inference time on DeepSportradar dataset (see Accuracy for details on dataset)
| Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
|--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------| -----------------------|
| [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 |
| [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 |
### Reference **MCU** memory footprint based on DeepSportradar dataset (see Accuracy for details on dataset)
|Model | Dataset | Format | Resolution | Series | Internal RAM | External RAM | Weights Flash | STEdgeAI Core version |
|----------|------------------|--------|-------------|------------------|------------------|---------------------|---------------|-------------------------|
| [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 |
| [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 |
### Reference **MCU** inference time on DeepSportradar dataset (see Accuracy for details on dataset)
| Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
|--------|------------------|--------|-------------|------------------|------------------|---------------------|-----------| -----------------------|
| [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 |
| [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 |
### Performance with DeepSportradar ReID dataset
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)
| Model | Format | Resolution | mAP | rank-1 accuracy |rank-5 accuracy |rank-10 accuracy |
|-------|--------|------------|----------------|-----------------|----------------|-----------------|
| [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 % |
| [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 % |
## Retraining and Integration in a simple example:
Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
# References
<a id="1">[1]</a>
The DeepSportradar Player Re-Identification Challenge (2023) [Online]. Available: https://github.com/DeepSportradar/player-reidentification-challenge. |