1st Place Solutions for UG2+ Challenge 2021 - (Semi-) supervised Face detection in the low light condition

To enhance the dark illumination of low light images, we employ the MSRCR [8] to it, which achieves simultaneous dynamic range compression/color consistency/lightness rendition. The enhanced image is shown in Figure 3. Besides, another data-driven brightness restoration method [6] is also used, which formulating light enhancement as a task of image-specific curve estimation with a deep network. The brightness restoration result is shown in Figure 5.

Additionally, the saliency map $R_{saliency}$ [7] extracted from the enhanced low light images $R_{msrcr}$, has been fused on $R_{msrcr}$ to suppress the false negative result, the fusion result $R_{saliency\_enhanced}$ is getting with:

where the $\alpha$ is set 0.3 in our work, and the result is shown in Figure 4

Different with using WIDERFACE and UFDD as pre-train sets, we merge them with pre-processed DARKFACE as a whole to build a more robust detector. Considering of the domain gap between pre-processed DARKFACE samples with the normal images (WIDERFACE, UFDD) that states in [14], we transfer WIDERFACE and UFDD to a more closer domain of processed DARKFACE set firstly. There are two different ways to achieve it, the traditional one is darkening the normal images, adding noise, and then processing it with MSRCR [8], the result is shown as Figure 6. Another method like [14], using the Pix2Pix network to synthesis noise, is shown in Figure 7. Based on the above Low Light Enhancement and Domain Transfer methods, we can obtain training samples with more closer domain that consist of low light enhanced images and domain transferred normal images.

We build a low light face detector based on the two-stage detection frameworks, including Cascade R-CNN [3], DetectoRS [11]. Using Cascade R-CNN as an example to describe the details, and the whole framework is shown in Figure 2.

Dataset Split Firstly, we split the DARKFACE set into several groups according to the number of faces in each image, and then randomly choose 10% samples of each group as the validation part, and the rest 90% data as the training part. We use the data augmentation method that described in section 2.1 to pre-process the DARKFACE samples. We also add WIDERFACE and UFDD dataset into our training set, which are pre-processed by the methods that described in section 2.2.

Training Strategy We apply the multi-scale training tricks with resizing samples range from [2160,1440] to [4320,2880], and apply random crop with size [1000,800] on it. Additionally, we use the popular image augment tools [2] to process our training samples online, including random flip and random lightness, color jittering, and several filtering methods, etc. Using AdamW optimizer with 0.0001 as initial learning rate, linear decay in 27 and 33 epoch with total 36 epoch, and the weight decay is 0.05.

Model Refinement Feature representation is always the key point to the object detection task. The backbone is very important to the feature representation ability of networks. Thus, we adopt two series powerful architecture Swin-Transformer [9] and ResNet [15] as our backbones. Besides, we also adapt PAFPN [13] to replace FPN in Cascade R-CNN. After analyzing the distribution of face size of the DARKFACE set, it is noticed that faces with small size is dominated which is illustrated in Figure 8. Thus we set more small anchors to capture more tiny faces. We also add attention modules such as GCnet [4] in the backbone to obtain more powerful representations. RoI-align module is also adopt in order to predict more precise bounding boxes.

Model Ensemble Finally, we trained Cascade R-CNN and DetectorRS with various backbones such as Swin-large, Swin-base, ResNet50 to obtain better diversity result of detectors. The performance of the models above that are trained on whole set can be found in the Table 1. We use Weighted Boxes Fusion (WBF) [12] and Test Time Augmentation(TTA) method for combining the predictions of our detectors, and the Soft-NMS[1] is employed before the model ensemble process.