MangaNet: Object Detection for Manga with Deep Neural Networks

Overview

Developed as a Stanford CS 231N project, MangaNet conducts the first controlled comparison of three fundamentally different detection paradigms — two-stage Faster R-CNN, anchor-based single-stage RetinaNet, and grid-based single-stage YOLOv3 — on the Manga109 benchmark (109 titles, 21,142 annotated pages, 500,000+ bounding boxes) for detecting 4 manga-specific object classes: body, face, frame, and text. Faster R-CNN with ResNet-50 + Feature Pyramid Network achieves a state-of-the-art COCO-standard mAP@[0.5:0.95] of 71.0, surpassing the prior SSD300-fork baseline (Ogawa et al., 2018). Beyond the detection benchmark, the project explores Neural Style Transfer as domain augmentation — applying artistic style transfer to manga pages and fine-tuning the detector on stylistically altered images to bridge the visual gap across the 109 manga titles’ diverse artistic styles. All models use ImageNet/COCO-pretrained backbones with full fine-tuning, custom imgaug-based augmentation pipelines with synchronized bounding box transformation, and COCO-standard evaluation via pycocotools.

Four Detection Targets

Three Detection Architectures

Faster R-CNN (Two-Stage) — Best Performer

Backbone: ResNet-50 + FPN (Feature Pyramid Network producing P2–P5 at 256 channels each). Stage 1: Region Proposal Network with 15 anchors per spatial location (5 scales × 3 aspect ratios, scales {32, 64, 128, 256, 512}px). Stage 2: RoI Align (7×7 pooling) → two FC layers (1024 units) → FastRCNNPredictor (5 classes: 4 object + 1 background). Loss: cross-entropy (classification) + Smooth L1 (box regression).

Initialized from full COCO-pretrained weights (not just backbone), then the classification head is replaced for 5-class output — all parameters remain trainable for full fine-tuning:

model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
in_features = model.roi_heads.box_predictor.cls_score.in_features
model.roi_heads.box_predictor = FastRCNNPredictor(in_features, num_classes=5)

RetinaNet (Single-Stage, Focal Loss)

Backbone: ResNet-50 + FPN (P3–P7, strides 8–128). Classification subnet: 4 Conv layers (256 channels, 3×3, ReLU) → K×A outputs. Box regression subnet: 4 Conv layers → 4×A outputs. Focal Loss: FL(pₜ) = −αₜ(1−pₜ)^γ · log(pₜ) with γ=2.0, α=0.25 to handle the extreme foreground-background class imbalance in dense manga pages. Initialized from ImageNet-pretrained backbone only — FPN and detection heads trained from scratch.

YOLOv3 (Single-Stage, Grid-Based)

Backbone: Darknet-53 (53 convolutional layers with residual connections). Multi-scale detection: 3 heads at 13×13 (large), 26×26 (medium), 52×52 (small) grids, 3 anchors per scale (9 total). Input: Fixed 416×416 pixels. Per-cell output: (bx, by, bw, bh, objectness, 4_class_probs) × 3 anchors. Pretrained from darknet53.conv.74 (first 74 layers on ImageNet), detection heads trained from scratch.

Dataset: Manga109

Data split (stratified via sklearn.train_test_split, random_state=0): 80% train / 10% validation / 10% test. Pages with zero annotations removed (“condensed” dataset). Degenerate bounding boxes (xmin ≥ xmax or ymin ≥ ymax) discarded.

YOLO format conversion: Absolute pixel coordinates [xmin, ymin, xmax, ymax] → normalized center coordinates [cx/W, cy/H, w/W, h/H] with per-image .txt label files.

Data Augmentation Pipeline

Custom imgaug-based pipeline ensuring synchronized transformation of images and bounding boxes:

class ImgAug(object):
    def __call__(self, image, target):
        bbs = BoundingBoxesOnImage(bbox_list, shape=image.shape)
        image, bbs = self.augmentations(image=image, bounding_boxes=bbs)
        bbs = bbs.clip_out_of_image()  # Clip boxes that went out of bounds

Training augmentation (DefaultAug): Sharpen (0.0–0.1) → Affine (translate ±10%, scale 0.8–1.5) → Brightness (−60 to +40) → Hue shift (±10) → Horizontal flip (50%) → PadSquare (center-pad to 1:1 aspect ratio) → ToTensor.

Validation/Test: PadSquare → ToTensor only (no augmentation).

StrongAug (available for ablation): Adds Dropout (0–1%), rotation (±10°), wider hue range (±20).

Training Configuration

Checkpoint management: Per-epoch full state saves + separate best-model checkpoint tracking {epoch, model_state_dict, optimizer_state_dict, mAP}, updated only when validation mAP improves. Supports resumable training from either best or most recent checkpoint.

Neural Style Transfer Domain Augmentation

A novel experimental approach to bridge the visual gap across the 109 manga titles’ diverse artistic styles:

1. Style transfer: Apply NST using a 21-style pretrained network (PyTorch-Style-Transfer) to generate stylistically altered versions of all manga pages (content size 1024px)
2. Two-stage fine-tuning: Start from a Faster R-CNN already trained on original manga data → fine-tune on NST-augmented images with SGD (lr=0.01, 10× higher than original) and batch size 8
3. Coordinate alignment: NST images resized to 1654×1170 to preserve annotation coordinate mappings

This explores whether artistic style transfer can serve as a domain augmentation strategy, making the detector more robust to the extreme visual diversity across manga series.

Evaluation

Primary metric: COCO-standard mAP@[0.5:0.95] — averaged over 10 IoU thresholds (0.50, 0.55, …, 0.95), the most stringent standard metric. Evaluated via pycocotools COCO API.

Why Faster R-CNN prevails: The two-stage architecture’s explicit region proposal mechanism excels for manga because: (1) the RPN generates high-quality proposals for densely overlapping body/face/frame/text regions, (2) FPN with multi-scale anchors (32–512px) handles the extreme range from tiny faces to full-page frames, and (3) full COCO model pretraining provides better initialization for detection-specific components (RPN, RoI heads) compared to backbone-only pretraining.

YOLOv3 inference: Confidence threshold 0.5, NMS IoU threshold 0.4, with box rescaling from 416×416 back to original dimensions and color-coded visualization per class.

Demo

Tech Stack

Python (3.7+), PyTorch, TorchVision (Faster R-CNN, RetinaNet), PyTorch-YOLOv3 (eriklindernoren), pycocotools (COCO evaluation), manga109api (annotation parsing), imgaug (augmentation with bbox sync), Pandas, NumPy, Matplotlib, Google Colab