Hugging Face's logo

STMicroelectronics
/
CNN2D-ST-HandPosture-model

Model card Files Community
FBAGSTM commited on
Commit
9799a9e
·
verified ·
1 Parent(s): 64a5546

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +91 -5
README.md CHANGED
@@ -1,5 +1,91 @@
1
- ---
2
- license: other
3
- license_name: sla0081
4
- license_link: https://github.com/STMicroelectronics/stm32ai-modelzoo/hand_posture/LICENSE.md
5
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: sla0081
4
+ license_link: https://github.com/STMicroelectronics/stm32ai-modelzoo/hand_posture/LICENSE.md
5
+ ---
6
+ # CNN2D_ST_HandPosture model
7
+
8
+ ## **Use case** : `Hand posture recognition`
9
+
10
+ # Model description
11
+
12
+ CNN2D_ST_HandPosture is a network topology designed by ST Teams to solve basic Hand Posture recognition use cases based on ST multi-zone Time-of-Flight sensor data. It is a convolutional neural network based model before feeding the data to the fully-connected (Dense) layer. It uses the distance and signal per spad 8x8 data. This is a very light model with very small foot prints in terms of FLASH and RAM as well as computational requirements.
13
+
14
+ We recommend to use input size (8 x 8 x 2) but this network can support greater input size.
15
+
16
+ The only input required to the model is the input shape and the number of outputs.
17
+
18
+ In this folder you will find multiple copies of the CNN2D_ST_HandPosture model pretrained on a ST custom datasets - Please refer to th [stm32ai-modelzoo-services](https://github.com/STMicroelectronics/stm32ai-modelzoo-services) GitHub for more informations
19
+
20
+ ## Network information (for 8 hand postures)
21
+
22
+
23
+ | Network Information | Value |
24
+ |:-----------------------:|:---------------:|
25
+ | Framework | TensorFlow |
26
+ | Params | 2,752 |
27
+
28
+
29
+ ## Network inputs / outputs
30
+
31
+
32
+ For an Time of Flight frame resolution of 8x8 and P classes
33
+
34
+ | Input Shape | Description |
35
+ | :----:| :-----------: |
36
+ | (N, 8, 8, 2) | Batch ( 8 x 8 x 2 ) matrix of Time of Flight values (distance, signal per spad) for a 8x8 frame in FLOAT32.|
37
+
38
+ | Output Shape | Description |
39
+ | :----:| :-----------: |
40
+ | (N, P) | Batch of per-class confidence for P classes in FLOAT32|
41
+
42
+
43
+ ## Recommended platforms
44
+
45
+
46
+ | Platform | Supported | Recommended |
47
+ |:--------:|:---------:|:-----------:|
48
+ | STM32F4 | [x] | [x] |
49
+ | STM32L4 | [x] | [x] |
50
+ | STM32U5 | [x] | [] |
51
+
52
+
53
+ # Performances
54
+
55
+ ## Metrics
56
+
57
+ Measures are done with default STM32Cube.AI configuration with enabled input / output allocated option.
58
+
59
+
60
+ ### Reference memory footprint based on ST_VL53LxCX_handposture_dataset (see Accuracy for details on dataset)
61
+
62
+ | Model | Format | Input Shape | Series | Activation RAM (KiB) | Runtime RAM (KiB) | Weights Flash (KiB) | Code Flash (KiB) | Total RAM (KiB) | Total Flash (KiB) | STM32Cube.AI version |
63
+ |:-----------------:|:------:|:-----------:|:-------:|:--------------:|:-----------:|:-------------:|:----------:|:-----------:|:-----------:|:---------------------:|
64
+ | [CNN2D_ST_HandPosture](https://github.com/STMicroelectronics/stm32ai-modelzoo/ST_pretrainedmodel_custom_dataset/ST_VL53L8CX_handposture_dataset/CNN2D_ST_HandPosture_8classes/CNN2D_ST_HandPosture_8classes.h5) | FLOAT32 | 8 x 8 x 2 | STM32F4 | 1.07 | 2.08 | 10.75 | 14.37 | 3.15 | 25.12 | 10.0.0 |
65
+ | [CNN2D_ST_HandPosture](https://github.com/STMicroelectronics/stm32ai-modelzoo/ST_pretrainedmodel_custom_dataset/ST_VL53L5CX_handposture_dataset/CNN2D_ST_HandPosture_8classes/CNN2D_ST_HandPosture_8classes.h5) | FLOAT32 | 8 x 8 x 2 | STM32F4 | 1.07 | 2.08 | 10.75 | 14.37 | 3.15 | 25.12 | 10.0.0 |
66
+
67
+
68
+ ### Reference inference time based on ST_VL53LxCX_handposture_dataset (see Accuracy for details on dataset)
69
+
70
+
71
+ | Model | Format | Resolution | Board | Frequency | Inference time (ms) | STM32Cube.AI version |
72
+ |:-----------------:|:------:|:----------:|:----------------:|:-------------:|:-------------------:|:---------------------:|
73
+ | [CNN2D_ST_HandPosture](https://github.com/STMicroelectronics/stm32ai-modelzoo/ST_pretrainedmodel_custom_dataset/ST_VL53L8CX_handposture_dataset/CNN2D_ST_HandPosture_8classes/CNN2D_ST_HandPosture_8classes.h5) | FLOAT32 | 8 x 8 x 2 | STM32F401 | 84 MHz | 1.54 ms | 10.0.0 |
74
+ | [CNN2D_ST_HandPosture](https://github.com/STMicroelectronics/stm32ai-modelzoo/ST_pretrainedmodel_custom_dataset/ST_VL53L5CX_handposture_dataset/CNN2D_ST_HandPosture_8classes/CNN2D_ST_HandPosture_8classes.h5) | FLOAT32 | 8 x 8 x 2 | STM32F401 | 84 MHz | 1.53 ms | 10.0.0 |
75
+
76
+ ### Accuracy with ST_VL53LxCX_handposture_dataset
77
+
78
+ Number of classes: 8 [None, FlatHand, Like, Dislike, Fist, Love, BreakTime, CrossHands]. Training dataset number of frames: 3,031. Test dataset number of frames: 1146.
79
+
80
+
81
+ | Model | Format | Resolution | Accuracy |
82
+ |:-----------------:|:------:|:----------:|:----------------:|
83
+ | [CNN2D_ST_HandPosture](https://github.com/STMicroelectronics/stm32ai-modelzoo/ST_pretrainedmodel_custom_dataset/ST_VL53L8CX_handposture_dataset/CNN2D_ST_HandPosture_8classes/CNN2D_ST_HandPosture_8classes.h5) | FLOAT32 | 8 x 8 x 2 | 99.43 % |
84
+ | [CNN2D_ST_HandPosture](https://github.com/STMicroelectronics/stm32ai-modelzoo/ST_pretrainedmodel_custom_dataset/ST_VL53L5CX_handposture_dataset/CNN2D_ST_HandPosture_8classes/CNN2D_ST_HandPosture_8classes.h5) | FLOAT32 | 8 x 8 x 2 | 97.17 % |
85
+
86
+
87
+ ## Retraining and Integration in a simple example:
88
+
89
+ Please refer to the stm32ai-modelzoo-services GitHub [here](https://github.com/STMicroelectronics/stm32ai-modelzoo-services)
90
+
91
+