<style>
.change-line{
display: block;
height: 10px;
}
</style>

# **AIA Object Detection Course Final Project**

**Final Project - Person Re-ID (April, 2022)**

<div class="change-line"></div>

*This model will contain:*

* Object Detection Model ( Getting the person )
* Classification Model ( Classification of the ID )

Flow of the process:

<div class="change-line"></div>

## **Grading**
* <font color="lightgreen">Simple baseline: 0.4 mAP</font>
* <font color="blue">Medium baseline: 0.5 mAP</font>
* <font color="red">Strong baseline: 0.6 mAP</font>
<div class="change-line"></div>

## **Object Detection Model**
* YOLOR ( You Only Learn One Representation )
* Original data ( 723 images with their annotations )
* Additional data ( ~ 900 images with thier annotations from CrowdHuman dataset )
<div class="change-line"></div>
### Detected Sample:

<div class="change-line"></div>

## **Classification Model**
* ResNet50
* Original data ( 100 classes, 5 ~ 20 images for each class )
* Advanced: OSNet
<div class="change-line"></div>

**New method**

* We apply ResNet50, removing the softmax layer, which results in an output vector with length = 2048.
* When the model obtains a cropped image from YOLOR, it will output a vector and save it in a gallery. This will form a Nx2048 matrix where N is the number of ID. If there are multiple cropped images belonging to the same ID, the vector will be updated with their mean value for each row.
* In the inferencing phase, the cropped image (query) from YOLOR will be sent to ResNet50 and output a feature vector. This feature vector will be compared with the vectors in the gallery.
* After the comparison, it will output the most possible ID, which means the feature vector of the query is similar to the one in the gallery.

<font color="red">**~~The first version didn't perform well due to the amount of the data, which implies we have try more way to do it such as Few-shot Learning~~.**</font>

<div class="change-line"></div>

<div class="change-line"></div>


<div class="change-line"></div>

## **Integration**
### Detected Sample:

<div class="change-line"></div>

## **Result of submission version 1:**

<font color='red'>**The performance wasn't good, it may be caused by the small amount of the data for training classification model.**</font>

## **Result of submission version 2:**

<font color='red'>**With slight progress...**</font>
<div class="change-line"></div>

## **Result of submission version 3:**

<font color='red'>**With slight progress...**</font>
<div class="change-line"></div>

## **Result of submission version 4:**

<font color='red'>**Achieved and exceeded the Strong baseline!**</font>
<div class="change-line"></div>

## **Result of submission version 5:**

<font color='red'>**Outpeform more accuracy than before!!!**</font>
<div class="change-line"></div>

## **Progress**
- [x] Object Detection Model
- [x] Classification Model
- [x] Integration
- [x] Submission

## **Leader Board**
https://docs.google.com/spreadsheets/d/10JYsHlLrqKtN4UwYUwsoUFsyjGdNQK5jI0y6RZvbbU0/edit#gid=0

**4/16/2022**
<font color='red'>$\rightarrow$ 1st place</font>

**4/15/2022**
<font color='red'>$\rightarrow$ 1st place</font>

**4/14/2022**
<font color='red'>$\rightarrow$ 1st place</font>