https://github.com/x-tabdeveloping/visual-analytics-assignment1
First assignment for visual analytics course.
https://github.com/x-tabdeveloping/visual-analytics-assignment1
Last synced: about 1 year ago
JSON representation
First assignment for visual analytics course.
- Host: GitHub
- URL: https://github.com/x-tabdeveloping/visual-analytics-assignment1
- Owner: x-tabdeveloping
- License: mit
- Created: 2024-02-23T13:23:25.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2024-05-13T16:31:23.000Z (almost 2 years ago)
- Last Synced: 2025-02-08T17:14:40.015Z (about 1 year ago)
- Language: Python
- Size: 459 KB
- Stars: 0
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# visual-analytics-assignment1
First assignment for visual analytics course.
This assignment is oriented at image retrieval in the [17 Category Flower Dataset](https://www.robots.ox.ac.uk/~vgg/data/flowers/17/)
## Setup
After downloading the data all jpg files should be arranged in one folder.
In the examples I will use the following folder structure:
```
- data/
- jpg/
- files.txt
- image_0001.jpg
...
```
> Beware!! The data directory in the repo is only added for demonstration puposes and for including the example images in the README.
The code ignore the `files.txt` file and scans the directory for all images.
This choice was made so that the code is easily reusable in non-indexed image datasets.
Each image is assumed to have a particular ID, this is the stem of the image path.
E.g. for the file `image_0001.jpg` its ID would be `"image_0001"`.
These IDs will be later used for specifying which image to base the search on.
To run the code first install dependencies:
```bash
pip install -r requirements.txt
```
## Usage
The repository contains code for a command line interface that can search in the dataset based on one user-specified image.
The CLI can use multiple distance metrics and types of latent representations to achieve this.
### Color Histograms
To do image retrieval based on minmax normalized color histograms, run the following:
```bash
python3 src/hist_search.py data/jpg/image_0001.jpg -o "out/hist"
```
This will put a CSV file with the images closest to the target by Chi Square histogram distance in the `out/hist` folder.
```
- out/
- hist/
- image_0001.csv
```
These are the results I got:
||Image|Distance|
|-|-|-|
|0|
|0.0|
|1|
|190.13992491162105|
|2|
|190.2249241130487|
|3|
|190.62783760197846|
|4|
|191.69055452774253|
|5|
|191.8753821638015|
As we can see many of the images barely resemble the original, and I would deem the performance of this method not satisfactory.
The problem likely lies in the histograms being too sparse (`256*256*256` bins) and the exact same colors rarely occur in two different images.
Increasing the size of the bins would likely yield much better results.
### VGG16
To use VGG16 image embeddings and cosine distance to search in the images run this command:
```bash
python3 src/embedding_search.py data/jpg/image_0001.jpg -o "out/vgg16"
```
This will put a CSV file with the images closest to the target cosine distance in the `out/vgg16` folder.
```
- out/
- vgg16/
- image_0001.csv
```
| |Image|Distance|
|-|-|-|
|0||1.369352364832821e-08|
|1|
|0.00986799891034007|
|2|
|0.010539749814202692|
|3|
|0.010623027922563977|
|4|
|0.011068601062943717|
|5|
|0.011395322683994236|
### Parameters
| Argument | Description | Type | Default |
|-------------------------|----------------------------------------------------------------------------------------------|---------|-------------------|
| `query_image_path` | Path to query image. | | - |
| `-h`, `--help` | Show this help message and exit. | | |
| `-i IMAGES_PATH`,
`--images_path IMAGES_PATH` | Path to the source directory containing images. | str | `"data/jpg"` |
| `-o OUT_PATH`,
`--out_path OUT_PATH` | Path to the output directory to save results. | str | `"out"` |
| `-k TOP_K`,
`--top_k TOP_K` | Top K similar images to retrieve. | int | 5 |