# S3FD_ATSS_SAPD

**Repository Path**: godycc/S3FD_ATSS_SAPD

## Basic Information

- **Project Name**: S3FD_ATSS_SAPD
- **Description**: 人脸识别
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2020-05-25
- **Last Updated**: 2020-12-19

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## S³FD_ATSS_SAPD: Single Shot Scale-invariant Face Detector ##
Applying [Adaptive Training Sample Selection](https://arxiv.org/abs/1912.02424) and [Soft Anchor Point Detection](https://arxiv.org/abs/1911.12448) to S3FD 
based on [PyTorch Implementation of Single Shot Scale-invariant Face Detector](https://github.com/yxlijun/S3FD.pytorch)
### Description
To train hand and head dataset with S3FD, hand dataset is [Egohands Dataset](http://vision.soic.indiana.edu/projects/egohands/) and head dataset is [SCUT-HEAD](https://github.com/HCIILAB/SCUT-HEAD-Dataset-Release)

### Requirement
* pytorch 1.X.X 
* opencv 
* numpy 
* easydict
* torchvision

### Prepare data 
1. download WIDER face dataset,Egohands dataset and SCUT-HEAD
2. modify data/config.py according to your home directory
3. ``` python prepare_wider_data.py ```
4. ``` python prepare_hand_dataset.py ```

### Train
I selected face dataset for the training and evaluation of ATSS and SAPD applications.
``` 
python train.py --batch_size 4 --dataset face
``` 

### Implementation Details
1. Adaptive Training Sample Selection(ATSS)

* I applied ATSS after the original sample selection part of S3FD in [bbox_utils.py](https://github.com/tgisaturday/S3FD_ATSS_SAPD/blob/master/layers/bbox_utils.py) (line 193-223)

* Unlike the original ATSS algorithm starting from empty candidate set, I used result positive set<br>
from stage two of S3FD as starting candidate set. Other details follow the original ATSS algorithm.

2. Soft Anchor Point Detection(SAPD)

* I applied SAPD to the smoothed_L1_loss of S3FD in [multibox_loss.py](https://github.com/tgisaturday/S3FD_ATSS_SAPD/tree/master/layers/modules/multibox_loss.py) (line 109-107)

* Anchor_weight calculation for generalized centerness function is done in [bbox_utils.py](https://github.com/tgisaturday/S3FD_ATSS_SAPD/blob/master/layers/bbox_utils.py) (line 293)

* I first multiply anchor_weight to the result of smoothed_L1_loss and devide the total sum of loss<br> 
with the sum of anchor_weight in [multibox_loss.py](https://github.com/tgisaturday/S3FD_ATSS_SAPD/tree/master/layers/modules/multibox_loss.py) (line 112-114)

* I tried to preserve the main concept of the original SAPD while modifying the generalized centerness<br>
function to make it fit to the original regression loss of S3FD.


### Evalution
I trained each model with WIDER face dataset for 100k iterations.

1. Extract predictions on WIDER FACE 
```
python wider_test.py
```
2. Evaluate predictions using [WiderFace-Evaluation](https://github.com/wondervictor/WiderFace-Evaluation). 

* Prediction results will be saved in 'eval_tools/s3fd_val'. 
* Ground truth files are stored in 'eval_tools/wider_gt/'.

### Result
1. Test on WIDER FACE 
```
Easy AP    Baseline = 0.946  ATSS_only = 0.944  SAPD_only = 0.943  ATSS_SAPD = 0.944
    
Medium AP  Baseline = 0.932  ATSS_only = 0.931  SAPD_only = 0.930  ATSS_SAPD = 0.932
    
Hard AP    Baseline = 0.849  ATSS_only = 0.846  SAPD_only = 0.847  ATSS_SAPD = 0.851
```

### References
* [S³FD: Single Shot Scale-invariant Face Detector](https://arxiv.org/abs/1708.05237)
* [Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection](https://arxiv.org/abs/1912.02424)
* [Soft Anchor-Point Object Detection](https://arxiv.org/abs/1911.12448)
* [ssd.pytorch](https://github.com/amdegroot/ssd.pytorch)
* [S3FD.pytorch](https://github.com/yxlijun/S3FD.pytorch)
* [WiderFace-Evaluation](https://github.com/wondervictor/WiderFace-Evaluation).