Computer Pointer Controller is an application which use a gaze detection model to control the mouse pointer of you computer.
The position of mouse pointer will change by following the user's Gaze. The Gaze Estimation model is used to estimate the gaze of the user's eyes and then feed the result into pyautogui module to change the position of mouse pointer.
The pipline of application as shown below:
Recorded video of running the project: Demo video of project
1.Prerequisites
- Install Intel® Distribution of OpenVINO™ toolkit for Windows* 10 or you can choose install in Linux system.
- The
requirments.txtin project directory needs to be installed. Using command:pip3 install -r requirements.txt
2.Environment setup Initialize openVINO environment (command in cmd)
!!!Important!!!
cd C:\Program Files (x86)\IntelSWTools\openvino\bin\ && setupvars.bat3.Download the required model
-
Download the required models:
-
These can be downloaded using the
model downloader. -
cd to project directory and follow command below to download the models.
python "C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\open_model_zoo\tools\downloader\downloader.py" --name face-detection-adas-binary-0001
python "C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\open_model_zoo\tools\downloader\downloader.py" --name head-pose-estimation-adas-0001
python "C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\open_model_zoo\tools\downloader\downloader.py" --name landmarks-regression-retail-0009
python "C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\open_model_zoo\tools\downloader\downloader.py" --name gaze-estimation-adas-0002The source structure of project as showm below:
E:\Intel-AI\Computer-Pointer-Controller-OpenVINO>tree /a /f
Folder PATH listing for volume entertainment
Volume serial number is 00000060 0003:B93B
E:.
| .Instructions.md.swp
| README.md
| requirements.txt
|
+---bin
| .gitkeep
| demo.mp4
| pipeline.png
| show_app.PNG
|
+---intel
| +---face-detection-adas-binary-0001
| | \---FP32-INT1
| | face-detection-adas-binary-0001.bin
| | face-detection-adas-binary-0001.xml
| |
| +---gaze-estimation-adas-0002
| | +---FP16
| | | gaze-estimation-adas-0002.bin
| | | gaze-estimation-adas-0002.xml
| | |
| | +---FP16-INT8
| | | gaze-estimation-adas-0002.bin
| | | gaze-estimation-adas-0002.xml
| | |
| | \---FP32
| | gaze-estimation-adas-0002.bin
| | gaze-estimation-adas-0002.xml
| |
| +---head-pose-estimation-adas-0001
| | +---FP16
| | | head-pose-estimation-adas-0001.bin
| | | head-pose-estimation-adas-0001.xml
| | |
| | +---FP16-INT8
| | | head-pose-estimation-adas-0001.bin
| | | head-pose-estimation-adas-0001.xml
| | |
| | \---FP32
| | head-pose-estimation-adas-0001.bin
| | head-pose-estimation-adas-0001.xml
| |
| \---landmarks-regression-retail-0009
| +---FP16
| | landmarks-regression-retail-0009.bin
| | landmarks-regression-retail-0009.xml
| |
| +---FP16-INT8
| | landmarks-regression-retail-0009.bin
| | landmarks-regression-retail-0009.xml
| |
| \---FP32
| landmarks-regression-retail-0009.bin
| landmarks-regression-retail-0009.xml
|
\---src
| face_detection.py
| facial_landmarks_detection.py
| file_explain.md
| gaze_estimation.py
| head_pose_estimation.py
| input_feeder.py
| main.py
| model.py
| mouse_controller.py
| test.py
| testarg.py
|
\---__pycache__
face_detection.cpython-37.pyc
facial_landmarks_detection.cpython-37.pyc
gaze_estimation.cpython-37.pyc
head_pose_estimation.cpython-37.pyc
input_feeder.cpython-37.pyc
mouse_controller.cpython-37.pyc
1. cd to src folder first
2. Template of run the main.py
python main.py -fd <Path of xml file for face detection model> -fl <Path of xml file for facial landmarks detection model> -hp <Path of xml file for head pose estimation model> -ge <Path of xml file for gaze estimation model> -i <Path of input video file or enter cam for feeding input from webcam> -d <choose the device to run the model (default CPU)> -show <select the visualization: win fd fl hp ge crop> -pc <use perf_counts to display performance of layers on each model> -info <show related information such as infer time of model, FPS, No.frame in command window>
3. Examples of run the main.py:
- Running model with percision FP32 in CPU:
python main.py -fd "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml" -fl "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP32\landmarks-regression-retail-0009.xml" -hp "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP32\head-pose-estimation-adas-0001.xml" -ge "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP32\gaze-estimation-adas-0002.xml" -i "\Intel-AI\Computer-Pointer-Controller-OpenVINO\bin\demo.mp4" -d CPU -show win fd fl hp ge -info -pc- Running model with percision FP16 in CPU:
python main.py -fd "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml" -fl "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP16\landmarks-regression-retail-0009.xml" -hp "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP16\head-pose-estimation-adas-0001.xml" -ge "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP16\gaze-estimation-adas-0002.xml" -i "\Intel-AI\Computer-Pointer-Controller-OpenVINO\bin\demo.mp4" -d CPU -show win fd fl hp ge -info -pc- Running model with percision FP16-INT8 in CPU:
python main.py -fd "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml" -fl "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP16-INT8\landmarks-regression-retail-0009.xml" -hp "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP16-INT8\head-pose-estimation-adas-0001.xml" -ge "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP16-INT8\gaze-estimation-adas-0002.xml" -i "\Intel-AI\Computer-Pointer-Controller-OpenVINO\bin\demo.mp4" -d CPU -show win fd fl hp ge -info -pc- Running model with percision FP32 in CPU and input through webcam
python main.py -fd "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml" -fl "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP32\landmarks-regression-retail-0009.xml" -hp "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP32\head-pose-estimation-adas-0001.xml" -ge "\Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP32\gaze-estimation-adas-0002.xml" -i cam -d CPU -show win fd fl hp ge -info -pcTry command python main.py -h to get help for command line arguments of the application
E:\Intel-AI\Computer-Pointer-Controller-OpenVINO\src>python main.py -h
usage: main.py [-h] -fd FACE_DETECTION_PATH -fl FACIAL_LANDMARKS_PATH -hp
HEAD_POSE_PATH -ge GAZE_ESTIMATION_PATH -i INPUT [-pc]
[-show FLAG_VISUALIZATION [FLAG_VISUALIZATION ...]] [-info]
[-l CPU_EXTENSION] [-d DEVICE] [-pt PROB_THRESHOLD]
optional arguments:
-h, --help show this help message and exit
-fd FACE_DETECTION_PATH, --face_detection_path FACE_DETECTION_PATH
(required) Path to Face Detection Model .xml file.
-fl FACIAL_LANDMARKS_PATH, --facial_landmarks_path FACIAL_LANDMARKS_PATH
(required) Path to Facial Landmarks Detection Model
.xml file.
-hp HEAD_POSE_PATH, --head_pose_path HEAD_POSE_PATH
(required) Path to Head Pose Estimation Model .xml
file.
-ge GAZE_ESTIMATION_PATH, --gaze_estimation_path GAZE_ESTIMATION_PATH
(required) Path to Gaze Estimation Model .xml file.
-i INPUT, --input INPUT
(required) Path to Input File either image or video or
CAM (using camera).
-pc, --perf_counts Report performance countersPrint the real time takes
for each layer in the model
-show FLAG_VISUALIZATION [FLAG_VISUALIZATION ...], --flag_visualization FLAG_VISUALIZATION [FLAG_VISUALIZATION ...]
(optional) Visualize the selected model on the output
frame window. 'win': display the visualization window
(To visualize other model must with 'win'), 'fd': Face
Detection Model, 'fl': Facial Landmarks Detection
Model, 'crop': Cropped face with Landmarks
Detection,'hp': Head Pose Estimation Model, 'ge': Gaze
Estimation Model. For example, '--show win fd fl hp ge
crop' (Seperate each flag by space).
-info, --show_info Flag to show all the related information in command
window. No. of frame, Average and total infer time of
each models, Total frame, FPS...
-l CPU_EXTENSION, --cpu_extension CPU_EXTENSION
(optional) MKLDNN (CPU)-targeted custom
layers.Absolute path to cpu_extension if layers from
model are not supported on device.
-d DEVICE, --device DEVICE
(optional) Specify the target device to infer on: CPU,
GPU, FPGA or MYRIAD is acceptable. Sample will look
for a suitable plugin for device specified (CPU by
default)
-pt PROB_THRESHOLD, --prob_threshold PROB_THRESHOLD
(optional) Probability threshold for detections (0.6
by default)
- FP32, FP16, FP16-INT8 tested on my CPU: Intel(R) Core(TM)i7-3632QM CPU @ 2.20GHz
- Checked each model load time, average inference time, total load model, total inference time, FPS and model size.
## Head pose estimation model performance:
name layer_type exet_type status real_time, us
48/Output_0/Data__const123_const_nchw_OIhw8i8o_pool1 Reorder jit_uni_FP32 NOT_RUN 0
60/Output_0/Data__const146_const Const unknown NOT_RUN 0
60/Output_0/Data__const146_const_nchw_OIhw8i8o_pool3 Reorder jit_uni_FP32 NOT_RUN 0
70/Output_0/Data__const114_const Const unknown NOT_RUN 0
70/Output_0/Data__const114_const_nchw_Ohwi8o_conv1 Reorder jit_uni_FP32 NOT_RUN 0
80/Output_0/Data__const137_const Const unknown NOT_RUN 0
80/Output_0/Data__const137_const_nchw_OIhw8i8o_conv3 Reorder jit_uni_FP32 NOT_RUN 0
Add_/Fused_Add_ ScaleShift jit_sse42_FP32 EXECUTED 35
Constant_11352 Const unknown NOT_RUN 0
Constant_11352_nchw_OhIw8o4i_conv2 Reorder ref_any_I8 NOT_RUN 0
Constant_11353 Const unknown NOT_RUN 0
Constant_11353_nchw_OhIw8o4i_conv4 Reorder ref_any_I8 NOT_RUN 0
Constant_11354 Const unknown NOT_RUN 0
Constant_11354_nchw_OhIw8o4i_conv5 Reorder ref_any_I8 NOT_RUN 0
Constant_11355 Const unknown NOT_RUN 0
Constant_11355_nchw_OhIw8o4i_conv_fm Reorder ref_any_I8 NOT_RUN 0
Constant_11356 Const unknown NOT_RUN 0
Constant_11356_nchw_OhIw8o4i_angle_y Reorder ref_any_I8 NOT_RUN 0
Constant_11359 Const unknown NOT_RUN 0
Constant_11359_nchw_OhIw8o4i_angle_r/Fused_Add_ Reorder ref_any_I8 NOT_RUN 0
Constant_11362 Const unknown NOT_RUN 0
Constant_11362_nchw_OhIw8o4i_angle_p Reorder ref_any_I8 NOT_RUN 0
Constant_12941 Const unknown NOT_RUN 0
Constant_12942 Const unknown NOT_RUN 0
Constant_12943 Const unknown NOT_RUN 0
Constant_12944 Const unknown NOT_RUN 0
Constant_12945 Const unknown NOT_RUN 0
Constant_12946 Const unknown NOT_RUN 0
Constant_12947 Const unknown NOT_RUN 0
Constant_12948 Const unknown NOT_RUN 0
Constant_12949 Const unknown NOT_RUN 0
Constant_12951 Const unknown NOT_RUN 0
Constant_12952 Const unknown NOT_RUN 0
Constant_12954 Const unknown NOT_RUN 0
Constant_12955 Const unknown NOT_RUN 0
Constant_12957 Const unknown NOT_RUN 0
angle_p Convolution jit_sse42_I8 EXECUTED 13
angle_p/WithoutBiases/fq_input_0 FakeQuantize undef NOT_RUN 0
angle_p_fc FullyConnected jit_gemm_FP32 EXECUTED 3
angle_p_fc/WithoutBiases/1_port_transpose3315_const184_const Const unknown NOT_RUN 0
angle_p_fc/flatten_fc_input Reshape unknown_FP32 NOT_RUN 0
angle_p_fc/flatten_fc_input_ScaleShift_angle_p_fc ScaleShift jit_sse42_FP32 EXECUTED 5
angle_p_nhwc_nchw_angle_p_fc/flatten_fc_input Reorder jit_uni_FP32 EXECUTED 5
angle_r/Fused_Add_ Convolution jit_sse42_I8 EXECUTED 16
angle_r/Fused_Add__nhwc_nchw_angle_r_fc/flatten_fc_input Reorder jit_uni_FP32 EXECUTED 5
angle_r_fc FullyConnected jit_gemm_FP32 EXECUTED 4
angle_r_fc/WithoutBiases/1_port_transpose3323_const195_const Const unknown NOT_RUN 0
angle_r_fc/flatten_fc_input Reshape unknown_FP32 NOT_RUN 0
angle_r_fc/flatten_fc_input_ScaleShift_angle_r_fc ScaleShift jit_sse42_FP32 EXECUTED 6
angle_y Convolution jit_sse42_I8 EXECUTED 15
angle_y_fc FullyConnected jit_gemm_FP32 EXECUTED 49
angle_y_fc/WithoutBiases/1_port_transpose3319_const210_const Const unknown NOT_RUN 0
angle_y_fc/flatten_fc_input Reshape unknown_FP32 NOT_RUN 0
angle_y_fc/flatten_fc_input_ScaleShift_angle_y_fc ScaleShift jit_sse42_FP32 EXECUTED 6
angle_y_nhwc_nchw_angle_y_fc/flatten_fc_input Reorder jit_uni_FP32 EXECUTED 6
bn1/variance/Fused_Add_ ScaleShift undef NOT_RUN 0
bn3/variance/Fused_Add_ ScaleShift undef NOT_RUN 0
conv1 Convolution jit_sse42_FP32 EXECUTED 142
conv2 Convolution jit_sse42_I8 EXECUTED 145
conv2/WithoutBiases/fq_input_0 FakeQuantize undef NOT_RUN 0
conv2_nhwc_nChw8c_pool2 Reorder jit_uni_FP32 EXECUTED 23
conv3 Convolution jit_sse42_FP32 EXECUTED 594
conv4 Convolution jit_sse42_I8 EXECUTED 89
conv4/WithoutBiases/fq_input_0 FakeQuantize undef NOT_RUN 0
conv5 Convolution jit_sse42_I8 EXECUTED 90
conv_fm Convolution jit_sse42_I8 EXECUTED 120
conv_fm/WithoutBiases/fq_input_0 FakeQuantize undef NOT_RUN 0
data Input unknown NOT_RUN 0
out_angle_p_fc Output unknown_FP32 NOT_RUN 0
out_angle_r_fc Output unknown_FP32 NOT_RUN 0
out_angle_y_fc Output unknown_FP32 NOT_RUN 0
pool1 Convolution jit_sse42_FP32 EXECUTED 190
pool1_FP32_nChw8c_U8_nhwc_conv2 Reorder jit_uni_FP32 EXECUTED 19
pool2 Pooling jit_avx_FP32 EXECUTED 26
pool2_ScaleShift_conv3 ScaleShift jit_sse42_FP32 EXECUTED 8
pool3 Convolution jit_sse42_FP32 EXECUTED 189
pool3_FP32_nChw8c_U8_nhwc_conv4 Reorder jit_uni_FP32 EXECUTED 13
pool4 Pooling jit_sse42_I8 EXECUTED 7
pool4/fq_input_0 FakeQuantize undef NOT_RUN 0
relu_conv1 ReLU undef NOT_RUN 0
relu_conv2 ReLU undef NOT_RUN 0
relu_conv3 ReLU undef NOT_RUN 0
relu_conv4 ReLU undef NOT_RUN 0
relu_conv5 ReLU undef NOT_RUN 0
relu_p_1 ReLU undef NOT_RUN 0
relu_r_1 ReLU undef NOT_RUN 0
relu_y_1 ReLU undef NOT_RUN 0
Average inference time of HeadPoseEstimation model: 16.855223137035704 ms
INFO:root:## Face Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml
INFO:root:## Landmarks Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP32\landmarks-regression-retail-0009.xml
INFO:root:## Headpose Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP32\head-pose-estimation-adas-0001.xml
INFO:root:## Gaze Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP32\gaze-estimation-adas-0002.xml
INFO:root:
Input path exists: \Intel-AI\Computer-Pointer-Controller-OpenVINO\bin\demo.mp4
INFO:root:## Loaded Input Feeder
INFO:root:FaceDetection load time: 402.233 ms
INFO:root:FacialLandmarkDetection load time: 202.0 ms
INFO:root:HeadPoseEstimation load time: 178.005 ms
INFO:root:GazeEstimation load time: 211.995 ms
INFO:root:Total Load time: 994.233 ms
INFO:root:## All model successfully loaded!
INFO:root:## Start inference on frame!
ERROR:root:flag_return: False. Video has reach to the end...
INFO:root:Average inference time of FaceDetection model: 50.7061845165188 ms
INFO:root:Average inference time of FacialLandmarkDetection model: 1.6961178537142478 ms
INFO:root:Average inference time of HeadPoseEstimation model: 3.2278723635915982 ms
INFO:root:Average inference time of GazeEstimation model: 3.6383079270185053 ms
INFO:root:Total inference time: 38241.687 ms
INFO:root:Total frame: 59
INFO:root:FPS: 1.5428063385806183
ERROR:root:### Camera Stream or Video Stream has reach to the end...###
INFO:root:## Face Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml
INFO:root:## Landmarks Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP16\landmarks-regression-retail-0009.xml
INFO:root:## Headpose Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP16\head-pose-estimation-adas-0001.xml
INFO:root:## Gaze Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP16\gaze-estimation-adas-0002.xml
INFO:root:
Input path exists: \Intel-AI\Computer-Pointer-Controller-OpenVINO\bin\demo.mp4
INFO:root:## Loaded Input Feeder
INFO:root:FaceDetection load time: 378.925 ms
INFO:root:FacialLandmarkDetection load time: 187.5 ms
INFO:root:HeadPoseEstimation load time: 187.506 ms
INFO:root:GazeEstimation load time: 203.12 ms
INFO:root:Total Load time: 957.05 ms
INFO:root:## All model successfully loaded!
INFO:root:## Start inference on frame!
ERROR:root:flag_return: False. Video has reach to the end...
INFO:root:Average inference time of FaceDetection model: 46.96024070351811 ms
INFO:root:Average inference time of FacialLandmarkDetection model: 0.5190655336541644 ms
INFO:root:Average inference time of HeadPoseEstimation model: 2.277273242756472 ms
INFO:root:Average inference time of GazeEstimation model: 3.472477702771203 ms
INFO:root:Total inference time: 36004.365 ms
INFO:root:Total frame: 59
INFO:root:FPS: 1.6387068103544051
ERROR:root:### Camera Stream or Video Stream has reach to the end...###
INFO:root:## Face Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\face-detection-adas-binary-0001\FP32-INT1\face-detection-adas-binary-0001.xml
INFO:root:## Landmarks Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\landmarks-regression-retail-0009\FP16-INT8\landmarks-regression-retail-0009.xml
INFO:root:## Headpose Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\head-pose-estimation-adas-0001\FP16-INT8\head-pose-estimation-adas-0001.xml
INFO:root:## Gaze Model path is correct: \Intel-AI\Computer-Pointer-Controller-OpenVINO\intel\gaze-estimation-adas-0002\FP16-INT8\gaze-estimation-adas-0002.xml
INFO:root:
Input path exists: \Intel-AI\Computer-Pointer-Controller-OpenVINO\bin\demo.mp4
INFO:root:## Loaded Input Feeder
INFO:root:FaceDetection load time: 328.122 ms
INFO:root:FacialLandmarkDetection load time: 218.754 ms
INFO:root:HeadPoseEstimation load time: 312.502 ms
INFO:root:GazeEstimation load time: 365.017 ms
INFO:root:Total Load time: 1224.395 ms
INFO:root:## All model successfully loaded!
INFO:root:## Start inference on frame!
ERROR:root:flag_return: False. Video has reach to the end...
INFO:root:Average inference time of FaceDetection model: 46.5065099425235 ms
INFO:root:Average inference time of FacialLandmarkDetection model: 0.9639626842434124 ms
INFO:root:Average inference time of HeadPoseEstimation model: 2.110372155399646 ms
INFO:root:Average inference time of GazeEstimation model: 4.738435906878973 ms
INFO:root:Total inference time: 35699.297 ms
INFO:root:Total frame: 59
INFO:root:FPS: 1.6527073587495449
ERROR:root:### Camera Stream or Video Stream has reach to the end...###
| Model | Type | Size | Load time | Average inference time |
|---|---|---|---|---|
| face-detection-adas-binary-0001 | FP32-INT1 | 1.86 MB | 402.233 ms | 50.7061845165188 ms |
| FP32-INT1 | 1.86 MB | 378.925 ms | 46.96024070351811 ms | |
| FP32-INT1 | 1.86 MB | 328.122 ms | 46.5065099425235 ms | |
| landmarks-regression-retail-0009 | FP32 | 786 KB | 202.0 ms | 1.6961178537142478 ms |
| FP16 | 413 KB | 187.5 ms | 0.5190655336541644 ms | |
| FP16-INT8 | 314 KB | 218.754 ms | 0.9639626842434124 ms | |
| head-pose-estimation-adas-0001 | FP32 | 7.34 MB | 178.005 ms | 3.2278723635915982 ms |
| FP16 | 3.69 MB | 187.506 ms | 2.277273242756472 ms | |
| FP16-INT8 | 2.05 MB | 312.502 ms | 2.110372155399646 ms | |
| gaze-estimation-adas-0002 | FP32 | 7.24 MB | 211.995 ms | 3.6383079270185053 ms |
| FP16 | 3.69 MB | 203.12 ms | 3.472477702771203 ms | |
| FP16-INT8 | 2.09 MB | 365.017 ms | 4.738435906878973 ms |
| Total Model Load time | Total Inference Time | FPS | |
|---|---|---|---|
| FP32 | 994.233 ms | 38241.687 ms | 1.5428063385806183 |
| FP16 | 957.05 ms | 36004.365 ms | 1.6387068103544051 |
| FP16-INT8 | 1224.395 ms | 35699.297 ms | 1.6527073587495449 |
- For different precision, the model size decreases follow the order of FP32 > FP16 > FP16-INT8. The total inference time follows the same order in this case, the INT8 is faster than FP16 and FP32 is slower than FP32. Lower precision model uses less memory. However, remember lower precision of model also lose the accuracy of model.
- Memory Access of FP16 is half the size compared with FP32, which reduces memory usage of a neural network. FP16 data transfers are faster than FP32, which improves speed (TFLOPS) and performance.
- The Model load time and FPS of them are almost the same, but Total model load time of FP16 is much faster than others. For FPS INT-8 is better than others. Besides model only load only when model initializes.
- In order to achieving the most reasonable combination, we do not want too longer inference time also too low accuracy. Moreover, in some specific scenario such as low budget. We do not want to waste storage attempt to get very high accuracy. To achieve the balance, we need to consider the volume of sacrifice.
- I would choose FP16 as best option based on all results, for each model the load time, average inference time, total load time and total inference time in FP16 gives very good results.
- Use the VTune Amplifier to find hotspots in inference engine pipline.
- Build an inference pipeline for both video file and webcam feed as input. Allow the user to select their input option in the command line arguments.
- Benchmark the running times of different parts of the preprocessing and inference pipeline and let the user specify a CLI argument if they want to see the benchmark timing. Use the get_perf_counts API to print the time it takes for each layer in the model.
- If there will multiple face situation, the model only extracts one face to control the mouse pointer and ignore the other faces.
- If due to certain reason, model couldn't detect the face. It will continue process another frame untill it detects face or keyboard interrupt to exit the program.