An implementation of RetinaFace: Single-stage Dense Face Localisation in the Wild [1] by Pytorch. The model achieved 68.15% average precision in WIDER FACE [2] hard set (on validation set) where I made few changes in its anchors box generative methods, its multi-task loss function and RetinaFace architecture (will be discuss in section below).
Feature Pyramid. We used Feature pyramid [3] on Resnet18 as the backbone of the model in order to extract multi-scale of image features. On the Object Detection task, the problem of detecting small object was a remarkable hard. The Feature pyramid computed a feature hierarchy layer by layer and produces feature maps of different spatial resolutions (multiple levels) as output where the high-resolution with weak feature can combine with strong low-resolution feature. Therefore, Feature pyramid can extremely increase semantic context value.
Figure 1. Top: a top-down architecture with skip connections, where predictions are made on the finest level. Bottom: our model that has a similar structure but leverages it as a feature pyramid, with predictions made independently at all levels. [3]
Context Module. The original design of the Context Module was inspired by Single Stage Headless Detector [4] to enhance the model's contextual reasoning power for capturing tiny faces. My RetinaFace model only use 1 Context Module shared for all feature pyramid output (while in original model are 5 seperate Context Module).
Figure 2. SSH context module [4]
Anchor Settings. When I took a quick survey on WIDER FACE dataset [2], it's shocking to found that number of samples which have width and height in between [1, 25] take ~71% and the most common bounding box aspect ratio are 1:2 and 1:1. Therefore, I made a few change in the generative method of the anchor box in order to strength the capability of detecting smail face, detail in table below:
| Feature Pyramid | Stride | Anchor | Number of anchor on feature map |
|---|---|---|---|
| P2 (120×120×64) | 4 | 8, 10.08, 12.67 | 86400 |
| P3 (60×60×64) | 8 | 16, 20.16, 25.40 | 21600 |
| P4 (30×30×64) | 16 | 32, 40.32, 50.80 | 5400 |
| P5 (15×15×64) | 32 | 64, 80.63, 101.59 | 1350 |
| C6/P6 (8×8×64) | 64 | 128, 161.26, 203.19 | 384 |
The bounding box are varible size from 8×8 up to 203×203 (this can be a minus because the model is not able to detect face have size bigger than 203×203 in 480×480 image) and 75% anchors box were came from P2.
Note: In pratice, I once set number of anchor on P2 up to ~172k boxes. However, I noticed that the model was fail with huge number of false positive prediction.
Loss function. Coming soon.
Due to the limitation of resource, the training process was trained on Colab Free GPU (with 15GB Vram). I have to resize the image into 480×480 (to advoid OOM) and set number of output channel on each feature map to 64 while it's 256 in original model and 640×640 as input shape. I used SGD optimizer with momentum = 0.9, weight decay = 5e-4, batch-size is 8 and learning rate starts from 1e-3 and decrease 0.7 time after 5 epochs. The process terminated after 30 epochs and the result are shown below:
| Backbone | EASY SET | MEDIUM SET | HARD SET |
|---|---|---|---|
| MobileNet0.25 (original image scale) | 90.70% | 88.16% | 73.82% |
| MobileNet0.25 (same parameter with Mxnet) | 88.67% | 87.09% | 80.99% |
| Resnet151 (Mxnet) | 96.942% | 96.175% | 91.857% |
| Resnet18 (our) | 74.35% | 71.65% | 68.15% |
Notes: You can find our (pre-processing) dataset and see training experiment through Weight&Bias.
While my MobileNet.25 doesn't have any pre-train weight, I encourage to use a pre-trained Resnet (18, 34, 50, 152) by calling e.g. --model resnet18 when training
Train a RetinaFace on WIDER FACE by specifying batch-size, pretained weight, epoch, trigger download pre-processed dataset from Wandb database, ...
# train RetinaFace on WIDER FACE with resnet18 as backbone
$ python train.py --image_size 640 --batchsize 8 --epochs 5 --model resnet18 --download --device 0
--batchsize 16 --model resnet34
--batchsize 32 --model resnet50
[1] Jiankang Deng andothers. “Retinaface: Single-stage dense face localisation in the wild”.in: arXiv preprint arXiv:1905.00641 (2019).
[2] Shuo Yang andothers. “Wider face: A face detection benchmark”. in: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pages 5525–5533.
[3] Tsung-Yi Lin andothers. “Feature pyramid networks for object detection”. in: Proceedings of the IEEE conference on computer vision and pattern recognition. 2017, pages 2117–2125
[4] Mahyar Najibi andothers. “Ssh: Single stage headless face detector”. in: Proceedings of the IEEE international conference on computer vision. 2017, pages 4875–4884.