ASFD: Automatic and Scalable Face Detector

Abstract

Along with current multi-scale based detectors, Feature Aggregation and Enhancement (FAE) modules have shown superior performance gains for cutting-edge object detection. However, these hand-crafted FAE modules show inconsistent improvements on face detection, which is mainly due to the significant distribution difference between its training and applying corpus, i.e. COCO vs. WIDER Face. To tackle this problem, we essentially analyse the effect of data distribution, and consequently propose to search an effective FAE architecture, termed AutoFAE by a differentiable architecture search, which outperforms all existing FAE modules in face detection with a considerable margin. Upon the found Auto-FAE and existing backbones, a supernet is further built and trained, which automatically obtains a family of detectors under the different complexity constraints. Extensive experiments conducted on popular benchmarks, i.e. WIDER Face and FDDB, demonstrate the state-of-the-art performance-efficiency trade-off for the proposed automatic and scalable face detector (ASFD) family. In particular, our strong ASFD-D6 outperforms the best competitor with AP 96.7/96.2/92.1 on WIDER Face test, and the lightweight ASFD-D0 costs about 3.1 ms, i.e. more than 320 FPS, on the V100 GPU with VGA-resolution images. CCS CONCEPTS• Computing methodologies → Object detection.

LGSNet: A Two-Stream Network for Micro- and Macro-Expression Spotting With Background Modeling

“…Video-level Distribution Consistency Strategy. Contrary to the general TAL [49], [50], [51], [52], expression spotting heavily relies on the sample distribution of different classes [10], [11]. We observe that the distribution of MEs in long, untrimmed face videos is sparser than that of MaEs on the CAS(ME) 2 [10] and SAMM-LV [11] datasets, because MEs are more challenging to evoke than MaEs [13].…”

“…Utilizing weak labels to train models has come a long way in computer vision such as semantic segmentation [44], [45], [46], object detection [47], [48], and temporal action localization (TAL) [21], [22], [23]. In contrast to the fullysupervised TAL [49], [50], [51], [52], the WTAL methods are free of extensive frame-level annotations and adopt video- [23], [53], [54], [55], [56] or point (key frame)-level [57], [58], [59], [60] labels during training. Since different videolevel WTAL approaches have different emphases, we can categorize them as foreground-only, background-assisted or pseudo-label-guided.…”

Online Learning State Evaluation Method Based on Face Detection and Head Pose Estimation

“…We conducted comparative experiments on RetinaFace [ 33 ], LFFD [ 13 ], CenterFace [ 20 ], ASFD [ 34 ], and the proposed GA-Face network on the FDDB dataset and WiderFace validation set. In Table 2 , GA-Face (ghost only) represents removing the SimAM module from the attention enhancement structure ( Figure 2 c) of GA-Face.…”

“…GA-Face is trained for 120 epochs on the WiderFace training set, the network weights are optimized using the Adam optimizer, a learning rate of 0.0005 is achieved, and the learning rate is reduced by 10 times when training reaches the 70th and 100th epochs. We conducted comparative experiments on RetinaFace [33], LFFD [13], CenterFace [20], ASFD [34], and the proposed GA-Face network on the FDDB dataset and WiderFace validation set. In Table 2, GA-Face (ghost only) represents removing the SimAM module from the attention enhancement structure (Figure 2c) of GA-Face.…”