YOLO-FaceV2: A Scale and Occlusion Aware Face Detector

• YOLO-FaceV2 fuses P2 features to retain pixel-level information for small faces and adds dilated-convolution RFE at P5 to enlarge the receptive field.The paper notes that adding P2 can slightly reduce accuracy on medium and large targets, which RFE is intended to ameliorate.Source: §1 · pp. 2–3
• The detector redesigns anchor sizes and proportions using the effective receptive field and face proportions to better suit bounding-box regression.The introduction explains that the theoretical receptive field can exceed the effective receptive field, making anchors based on it harder to regress.Source: §1 · p. 3
• The regression loss incorporates Normalized Wasserstein Distance to mitigate IoU’s sensitivity to small-face localization deviations.The introduction states that slight position deviations can sharply reduce IoU for small targets.Source: §1 · p. 3
• For occlusion, Repulsion Loss discourages predicted boxes from shifting toward other ground truths and reduces NMS-threshold sensitivity, while SEAM enhances face-feature learning when objects cause feature disappearance.The introduction distinguishes occlusion between faces from occlusion by other objects.Source: §1 · pp. 3–4
• Slide weights samples by IoU, uses the mean IoU of candidate positive samples as the positive-negative dividing line, and gives harder samples greater weight.This design addresses easy-sample dominance while reducing the hyperparameter-tuning burden of prior hard-sample methods.Source: §1 · pp. 2–4

Face Detection

• Early face detectors used cascades, landmark alignment, or angle correction, while later methods adapted generic detectors such as Faster R-CNN, SSD, and RetinaNet.Examples include CascadeCNN, MTCNN, PCN [3], CMS-RCNN [34], FaceBoxes [25], and SRN [35]; the paper notes early methods’ drawbacks, including slow detection and low accuracy.Source: §2 · p. 4

Scale-invariance

• Scale variation is addressed through scale-invariant features, fewer down-sampling layers, dilated convolution, denser anchors, multi-scale training, and fusion of deep semantics with shallow features.The discussion also cites SNIP [43] and TridentNet [12] as approaches to the multi-scale problem.Source: §2 · pp. 4–5

Occlusion problem

• Contextual features are a recurring approach to occlusion, used by SSH [37], FAN [13], PyramidBox [44], and RetinaFace [45] to strengthen information around partially visible faces.Occlusion can cause missed detections and low recall when face regions are invisible or boundaries blurred.Source: §2 · p. 5

Imbalance of easy and hard samples

• Sample imbalance methods include OHEM [46], Focal Loss [6], SRN [35], and FaceBoxes [25], but the paper notes their hyperparameters can be difficult to tune.The authors propose a sample-balance function with adaptive parameters.Source: §2 · p. 5

3.1 Network Architecture

• CSPDarknet53 uses RFE in P5, P2 features join the SPP/PAN neck, and an occlusion-focused SEAM branch follows the neck.The heads classify faces and regress their locations.Source: §3.1 · p. 5, Figure 1 · p. 6

3.2 Scale-Aware RFE Model

• RFE combines dilated-convolution branches with shared weights, residual connection, and a gathering-and-weighting layer to fuse multi-scale features and expand the receptive field.It replaces the Bottleneck in the P5 CSP block.Source: §3.2 · p. 6, Figure 2 · p. 7

3.3 Occlusion-Aware Repulsion Loss

• Repulsion Loss uses RepGT to push a predicted box away from the nearest other ground-truth face and RepBox to reduce overlap between predictions for different faces.RepBox aims to prevent NMS from suppressing one of two faces represented by overlapping predictions.Source: §3.3 · pp. 7–8

3.4 Occlusion-Aware Attention Network

• SEAM combines depthwise separable convolutions, pointwise convolutions, and channel-wise fully connected layers to emphasize face regions and use learned channel relationships to recover occluded features.Exponentially normalized attention is multiplied by the original features and is designed to be more tolerant to position error.Source: §3.4 · p. 8

3.5 Sample weighting function

• Slide weighting divides samples at the mean IoU threshold µ and assigns higher weights near the boundary to focus training on difficult samples.The authors identify boundary samples as having unclear classifications and relatively large losses.Source: §3.5 · p. 8, Figure 4 · p. 10

3.6 Anchor Design Strategy

• Anchors use a 1:1.2 width-to-height ratio and sizes based on each layer’s effective receptive field, which is smaller than its theoretical receptive field.The three detection heads are associated with anchor scales designed according to the P2, P3, and P4 strides.Source: §3.6 · p. 9, Figure 5 · p. 10

3.7 Normalized Gaussian Wasserstein Distance

• NWD Loss supplements IoU loss for small-target regression because its Gaussian-distribution similarity measure is less sensitive to target scale, while IoU remains for large targets.NWD measures similarity between predicted and real targets whether or not they overlap.Source: §3.7 · p. 10

• The experiments compare the proposed detector’s performance with other state-of-the-art face detectors.Source: §4 · p. 11
• Ablations assess the effectiveness of the proposed attention module.Source: §4 · p. 11
• Ablations assess the multi-scale fusion pyramid structure.Source: §4 · p. 11
• Ablations assess the proposed loss function design.Source: §4 · p. 11

4.1 Dataset

• WiderFace contains 32,203 images and more than 400,000 faces, split into 40% training, 10% verification, and 50% test data.The model was trained on the training set and evaluated on the validation and test sets.Source: §4.1 · p. 11
• The hard subset is the most challenging and is described as better reflecting face-detector effectiveness.The dataset is divided into easy, medium, and hard subsets.Source: §4.1 · p. 11

4.2 Training

• Training used SGD with learning rates of 1e-2 initially and 1e-3 finally, weight decay 5e-3, and momentum 0.8 for three warm-up epochs then 0.937.The NMS IoU was 0.5; training ran on a 1080ti with four CPU workers for 100 iterations at batch size 16.Source: §4.2 · p. 11

4.3.1 SEAM Block

• SEAM increases accuracy by 0.88, 0.82, and 1.06 on the easy, medium, and hard validation subsets, respectively.The attention block compensates for occluded-face response loss by strengthening responses to unobstructed faces.Source: §4.3.1 · p. 11, Table 2 · p. 12

4.3.2 Multi-scale feature fusion

• RFE increases accuracy by 0.5, 0.6, and 2.17, respectively, while adding P2 features raises the hard-subset result by 0.57.P2 fusion adds small-target information; RFE uses dilated convolutions with rates 1, 2, and 3 to improve long-range dependency.Source: §4.3.2 · p. 12, Table 2 · p. 12

4.3.3 Slide Loss

• Slide slightly improves results on the medium and hard subsets by focusing the model more on hard samples.Source: §4.3.3 · p. 12, Table 2 · p. 12

4.3.4 Anchor Design

• Designed anchors improve results by 0.24, 0.75, and 0.9 on the easy, medium, and hard datasets, respectively.Anchor ratios and sizes are designed using effective receptive field and face-shape characteristics to recall more small faces.Source: §4.3.4 · p. 12, Table 2 · p. 12

4.3.5 NWD Loss

• Replacing IoU with NWD alone does not improve results, so the study retains IoU and adjusts its proportion relative to NWD.The rationale is that IoU performs better for large and medium targets, while NWD can improve detection accuracy for small targets.Source: §4.3.5 · p. 12

4.3.6 Balance of RepGT and RepBox

• Repulsion Loss improves accuracy by 0.71, 0.63, and 0.5 on the easy, medium, and hard subsets, respectively.The study varies face-occlusion thresholds to adapt the loss to face detection.Source: Table 2 · p. 12, §4.3.6 · p. 13

4.4 Comparisons with Existing Face Detectors

• Table 4 compares YOLO-FaceV2 with existing face detectors on WiderFace validation, grouped by underlying detector family and using data from the dataset’s official website.The listed families include Fast R-CNN, SSD, and YOLO.Source: Table 4 · p. 13, §4.4 · p. 15
• Figure 6 shows precision-recall curves for YOLO-FaceV2 and competing detectors on the Easy, Medium, and Hard WiderFace validation subsets.Source: Figure 6 · p. 14, §4.4 · p. 15

• YOLO-FaceV2 achieves close to or exceeding SOTA performance on the WiderFace validation Easy and Medium subsets.Source: §5 · p. 15
• For varying face scales, it fuses the P2 layer into the feature pyramid, uses RFE to enhance receptive field, and applies NWD Loss for robust small-target detection.Source: §5 · p. 15
• Slide alleviates easy-hard sample imbalance, while SEAM and Repulsion Loss address occlusion; effective receptive field information guides anchor design.Source: §5 · p. 15