Seeing Small Faces from Robust Anchor's Perspective

1. Introduction

• Anchor-based detectors match faces to predefined boxes using IoU, then classify and regress matched anchors during training and inference.The approach builds on anchor-based deep CNN detectors used in current state-of-the-art face detection.Source: §1 · p. 1
• Faces smaller than 16 × 16 pixels have substantially lower recall than larger faces, and average maximum anchor IoU is positively correlated with recall across face scales.The authors suggest low initial overlap makes small faces harder to adjust to ground truth.Source: §1 · pp. 1–2, Figure 1 · p. 1
• Expected Max Overlapping (EMO) estimates a face’s expected maximum IoU with anchors under a location distribution, explaining why larger faces and denser anchors achieve higher overlap.The proposed anchors have higher face IoU than the baseline anchors.Source: §1 · p. 2
• The proposed design strategies reduce anchor stride, add anchors shifted from the canonical center, stochastically shift faces, and match low-overlap faces with multiple anchors.The strategies target higher face IoU without adding much network complexity.Source: §1 · p. 2
• The method achieves state-of-the-art performance on Wider Face, AFW, PASCAL Faces, and FDDB, with competitive runtime speed.It also significantly improves over the baseline anchor-based method.Source: §1 · p. 2

2. Related Work

• Earlier face detectors used hand-crafted features, including boosting with rigid templates and Deformable Parts Models, and optimized components disjointly.The cited work describes this disjoint optimization as making training less optimal.Source: §2 · p. 2
• Deep CNN face detectors improved performance through more discriminative features and end-to-end training.The passage reports that the performance gap between human and artificial face detectors had significantly closed.Source: §2 · p. 2
• Prior approaches to tiny faces use body context, multilevel image pyramids, or scale-equitable frameworks with new anchor matching strategies.This paper instead introduces a novel anchor design for finding more tiny faces.Source: §2 · p. 2

3.1. Overview of Anchor-Based Detector

• Anchor-based detectors match predefined boxes to ground-truth boxes by IoU, then classify and regress the matched anchors.An anchor is assigned if its IoU is highest or exceeds a threshold; anchors below a lower threshold for all boxes are labeled background.Source: §3.1 · p. 3

3.2. Anchor Setup and Matching

• The standard anchor set combines predefined box scales and aspect ratios with regularly spaced locations, with anchor stride equal to feature stride.At each location, the setup uses the Cartesian product of scales, ratios, and locations.Source: §3.2 · p. 3, §3.1 · p. 3

3.3. Computing the EMO Score

• For the Wider Face setting, the analysis uses six square anchor scales and assumes face centers are uniformly distributed across the image.The analysis exploits periodicity by considering one region where a face center lies among four anchor centers, and evaluates a face whose scale matches an anchor.Source: §3.3 · p. 4
• EMO is the expected IoU between a face and its best-matching anchor over possible face locations, and smaller anchor stride yields higher EMO for a fixed face size.The authors state that increasing EMO statistically increases the average maximum IoU across faces.Source: §3.3 · p. 4

4. Strategies of New Anchor Design

• The overall design seeks higher IoU by reducing anchor stride and face-to-anchor center distance, with multiple strategies targeting tiny faces.The section also proposes matching low-overlapped faces with multiple anchors.Source: §4 · p. 4

4.1. Stride Reduction with Enlarged Feature Maps

• Three architectures enlarge feature maps to reduce anchor stride; dilated convolutions achieve this without additional parameters while preserving receptive-field size.Bilinear upsampling and skip connections are the other described designs.Source: §4.1 · pp. 4–5, Figure 5 · p. 5

4.2. Extra Shifted Anchors

• Extra supportive anchors lower anchor stride below feature stride while sharing the center anchors’ feature representation.The anchors are regularly tiled, and adding small shifted anchors can further improve average IoU while saving computation relative to adding large anchors.Source: §4.2 · p. 5, Figure 6 · p. 5

4.1. Stride Reduction with Enlarged Feature Maps

• Halving anchor stride substantially increases average IoU for tiny faces compared with the original design (Figure 7).The comparison uses the proposed stride-reduction techniques.Source: Figure 7 · p. 5, §4.1 · p. 5

4.3. Face Shift Jittering

• Randomly shifting each training image changes face locations and increases the chance of higher face–anchor overlap.Offsets are sampled discretely from 0 to sA/2 −1, reflecting the stated overlap-pattern period.Source: §4.3 · pp. 5–6

4.4. Hard Face Compensation

• Because some very tiny faces remain poorly matched, the compensation strategy labels the top N overlapping anchors positive for faces below the overlap threshold.Anchors above the threshold are first labeled positive under the existing matching strategy.Source: §4 · p. 4, §4.4 · p. 6

5.1. Ablation Study

• Up to 7.4%: Enlarged-feature-map architectures improve Hard-subset performance over baseline; DC performs best across all three subsets without extra parameters.BUS improves tiny-face detection over BU but produces more false positives, attributed to its shallower, less discriminative res3 features.Source: §5.1 · p. 6
• Three added scale-16 anchors give the best performance; adding more anchors does not always improve AP because each anchor is trained from fewer matched faces.The shifted-anchor settings add one or three anchors to reduce stride, starting with scale 16.Source: §5.1 · pp. 6–7
• Shift jittering raises Hard-face AP by 0.9% with DC and by 0.3% with DC plus three scale-16 anchors.The shifts are applied randomly to faces at each training iteration.Source: §5.1 · p. 7
• Hard-face compensation completes the final model: assigning hard faces to N=5 anchors is a proper choice, since smaller N lowers recall and larger N yields more false positives.The final detector is denoted “+DC+16x3+SJ+HC(5)”.Source: §5.1 · p. 7
• At 600x1000 input size, the single-scale detector beats multiscale-trained SSH by 7.1% on Hard faces.Testing size affects precision, especially for tiny faces; image pyramids improve baseline performance, and the proposed strategies add further gains. Ablations use original images without rescaling.Source: §5.1 · pp. 6–7, Table 2 · p. 7

5.2. Evaluation on Common Benchmarks

• The final detector achieves consistent state-of-the-art performance on Wider Face, AFW, PASCAL Faces, and FDDB.The supplied benchmark figures show precision–recall curves on Wider Face and comparisons on AFW, PASCAL Faces, and FDDB.Source: §5.2 · p. 7, Figure 8 · p. 8, Figure 9 · p. 8

5.3. Runtime Speed

• The detector forms an upper envelope of other detectors in average precision versus inference time on Wider Face validation.The comparison varies image size; all detectors are evaluated on one NVIDIA Titan X GPU with batch size 1.Source: §5.3 · p. 8, Figure 10 · p. 8

6. Conclusion

• Low face–anchor overlap is identified as a major obstacle to tiny-face detection, motivating EMO analysis and anchor redesign that outperforms the baseline and reaches state of the art on four datasets.The reported datasets are Wider Face, AFW, PASCAL Faces, and FDDB.Source: §6 · p. 8

• Anchor stride reduction and extra shifted anchors both help recall more tiny faces.Source: S1. Recall Rate Comparison · p. 11
• The new anchor design improves average IoU by a large margin for small faces.Source: S1. Recall Rate Comparison · p. 11
• Anchors closer to faces are easier to regress to face boxes.Source: S1. Recall Rate Comparison · p. 11

S2. Dilated Convolution

• Dilated convolution yields the highest AP scores among the three architectures for reducing anchor stride.The comparison includes bilinear upsampling and bilinear upsampling with skip connection.Source: S2. Dilated Convolution · p. 11
• The figure contrasts normal and dilated convolution using a toy 1D visualization of 3 × 3 filters.It depicts stride-2 convolution followed by stride-1 convolution versus removing stride 2 and dilating the filter.Source: Figure S2 · p. 11, S2. Dilated Convolution · p. 11
• Unlike bilinear upsampling, dilated convolution fills holes layer by layer through transformations from shallow to deep layers.Source: S2. Dilated Convolution · p. 11
• Bilinear upsampling fills holes with interpolated values, while skip connections add shallower-layer features that are less discriminative than deep-layer features.Source: S2. Dilated Convolution · p. 11

S3. More Precision-Recall Curves

• The precision-recall curves and AP scores are reported for the Wider Face validation set.Source: S3. More Precision-Recall Curves · p. 11, Figure S3 · p. 12
• The detector consistently achieves state-of-the-art performance on Wider Face validation across Easy, Medium, and Hard levels.Source: S3. More Precision-Recall Curves · p. 12, Figure S3 · p. 12
• The comparison covers the Easy, Medium, and Hard validation splits.Source: S3. More Precision-Recall Curves · p. 12, Figure S3 · p. 12

S4. Qualitative Results

• The qualitative examples span Wider Face, AFW, PASCAL Faces, FDDB, and Internet images.Source: S4. Qualitative Results · p. 12, Figure S4 · p. 13, Figure S5 · p. 13
• Detected faces are annotated with green boxes and confidence scores, and readers are advised to zoom in on tiny-face detections.Source: S4. Qualitative Results · p. 12
• Wider Face examples include scale, pose, occlusion, expression, makeup, and illumination challenges.Source: S4. Qualitative Results · p. 12, Figure S4 · p. 13