ITMO-mobileCV
This repository contains all laboratory works for the ITMO University mobile systems for computer vision course.
Lab 1 (passed)
Captures a camera image and displays it with the area of interest in the center (60x60). When the image in that area is completely red, green or blue, its border is changed to the corresponding color. This program uses HSV color model.
Usage: python var1.py [-h <int>|--hue=<int>] [-s <int,int>|--saturation=<int,int>] [-v <int,int>|--value=<int,int>] [--fps=<int>]
Params:
-h or --hue - Sensitivity of Hue param in HSV model (optional)
-s or --saturation - Min and max values of Saturation (optional)
-v or --value - Min and max values of third param in HSV model (optional)
--fps - Frames per second to capture (optional)
Defaults:
Hue sensitivity: 20
Saturation: 70..255
Value: 50..255
Lab 2 (passed)
2 implementations of mean filter with hardcoded 3x3 area for each pixel. void mean_filter(Mat, Mat, int, int) function uses vectors to proceed image filtering. void mean_filter_neon(uint8_t*, uint8_t*, int, int) takes an advantage of parallel computing using Arm NEON command system.
Compile: make
Usage: ./var1 <imagepath>
Output: mean.png and neon.png as the result of each function
Testing:
This code was tested on Jetson Nano platform. The results are listed below in the table.
| Code optimization |
Image resolution |
Mean, us | Mean NEON, us |
|---|---|---|---|
| 0 | SD | 186677 | 31728 |
| 1 | SD | 20457 | 5935 |
| 2 | SD | 19770 | 5429 |
| 3 | SD | 10169 | 4776 |
| 0 | HD | 211443 | 35402 |
| 1 | HD | 21619 | 6256 |
| 2 | HD | 22037 | 6254 |
| 3 | HD | 11333 | 5433 |
| 0 | Full HD | 1258991 | 209369 |
| 1 | Full HD | 130749 | 38703 |
| 2 | Full HD | 126249 | 37714 |
| 3 | Full HD | 65100 | 33504 |
| 0 | Ultra HD | 5007538 | 854234 |
| 1 | Ultra HD | 528413 | 172613 |
| 2 | Ultra HD | 508814 | 168508 |
| 3 | Ultra HD | 258277 | 152980 |
Lab 3 (passed)
Classifies input images via pretrained AlexNet neural network using PyTorch and torchvision.
Usage: python lab3.py [--trt] images...
Output: ./output/<image_name> with class ID and label at the bottom of the image.
Params:
--trt - Enables TensorRT optimization (optional)
Testing:
This code was tested on Jetson Nano platform. The results are listed below in the table.
| Name | Weights load, sec | Images processing, sec | Max memory allocated, byte |
|---|---|---|---|
| Without TRT | 20.54 | 26.13 | 257948160 |
| TRT | 285 | 3.5 | 606208 |
| TRT (saved state) | 40 | 3.5 | 606208 |
Lab 4 (passed)
Captures a camera image (or video) and locates faces on it. When the trained KNN classifier recognizes the face, a screenshot is captured and a new record with name and time is added to CSV file. This program uses face_recognition and scikit-learn libraries to locate and classify people.
Usage: python lab4.py [-d <dir>|--data=<dir>] [-v <path>|--video=<path>] [-c <path>|--clf==<path>]
Output: ./output/attendance.csv and screenshots of captured people at first time.
Params:
-d or --data - Train data directory (default: ./data)
-v or --video - Video to capture (default: camera capture)
-c or --clf - Load classificator state from file or create new one (optional)