bleakie / retinadetector

基于RetinaFace的目标检测方法，适用于人脸、缺陷、小目标、行人等

RetinaDetection Object Detector

Introduction

RetinaDetector是基于RetinaFace修改的检测方法，原论文is a practical single-stage SOTA face detector which is initially described in arXiv technical report

Data

0. Organise the dataset directory as follows:

  data/retinaface/
    train/
      images/
      label.txt
    val/
      images/
      label.txt
    test/
      images/
      label.txt

Install

1. Install MXNet with GPU support.
2. Install Deformable Convolution V2 operator from Deformable-ConvNets if you use the DCN based backbone.
3. Type make to build cxx tools.

Training

Please check train.py for training.

1. Copy rcnn/sample_config.py to rcnn/config.py

为了获得更好的训练效果，可针对性的修改一些参数，如下：

config.TRAIN.MIN_BOX_SIZE = 10 #最小bbox
config.FACE_LANDMARK = False #使用landmark
config.USE_BLUR = False
config.BBOX_MASK_THRESH = 0
config.COLOR_MODE = 2 #增强
config.COLOR_JITTERING = 0.125

无效人脸的过滤，如下：

if (x2 - x1) < config.TRAIN.MIN_BOX_SIZE or (y2 - y1) < config.TRAIN.MIN_BOX_SIZE:
   continue
if self._split.startswith('train'):
   blur[ix] = values[19]
   if blur[ix] < 0.25:
      continue
if config.BBOX_MASK_THRESH > 0:
   if (x2 - x1) < config.BBOX_MASK_THRESH or (y2 - y1) < config.BBOX_MASK_THRESH:
      boxes_mask.append(np.array([x1, y1, x2, y2], np.float))
      continue
   if self._split.startswith('train'):
      if blur[ix] < 0.35:
         boxes_mask.append(np.array([x1, y1, x2, y2], np.float))
         continue

2. Download pretrained models and put them into model/.

   ImageNet ResNet50 (baidu cloud and dropbox).

   ImageNet ResNet152 (baidu cloud and dropbox).

3. Start training with sh train_model.sh.
   Before training, you can check the resnet network configuration (e.g. pretrained model path, anchor setting and learning rate policy etc..) in rcnn/config.py.

Testing

Please check test.py for testing.

Result

1. 缺陷检测

2. 人脸检测+人脸对齐

Models

人脸检测模型，比原版误检更低，角度较大和模糊超过0.6的face会自动忽略，更适合人脸识别的应用：click here.

ToDo

由于缺陷检测数据涉及私密性，缺陷检测的模型暂时不会释放

References

@inproceedings{[email protected],
year={2019}
}

retinadetector's Issues

False Face Predictions?

I run your test.py file with your given face model
It is predicting too much false faces
Detecting blur faces too

ModuleNotFoundError: No module named 'symbol_common'

File "C:\Users\Administrator\Desktop\RetinaDetector-master\rcnn\symbol\symbol_mnet.py", line 9, in
from symbol_common import get_sym_train
ModuleNotFoundError: No module named 'symbol_common'

请问能说下label.txt里面的格式吗？采用绝对坐标还是相对坐标？

请问能说下label.txt里面的格式吗？采用绝对坐标还是相对坐标？

About trian dataset ground truth

Can you share a specific labeled example in your dataset?

关于比原版误检更低

@bleakie Hi，想问下你主要改进了什么使得RetinaDetector比原版误检更低？角度较大和模糊超过0.6的face自动忽略？无效人脸过滤？

Python version

Hi Bleakia,
In this code which python version (2.7 or 3.6) are you using?
Best regards,
PeterPham

关于数据的标注问题

请问可以分享一小段您的数据集的具体的标注方式嘛？

请问能识别4~10个像素大小的目标吗？

小白问个问题，请问config.TRAIN.MIN_BOX_SIZE = 10 #最小bbox，是指框的最短边长度吗？还是就是指bbox的面积是10？能识别4~10个像素大小的目标吗？

请问RetinaDetector与YOLO4, YOLO5相比，效果如何？

冒昧请教下，RetinaDetector与YOLO4, YOLO5相比，效果如何？在COCO上的表现如何？对小目标的检测效果相比如何？有什么优势和不足？

请问人脸检测数据集如何清洗的，可否提供label呀

请问该项目试用于多类别嘛？