VLR-CVC/ComicsPAP

Comics: Pick-A-Panel

Updated val and test on 25/02/2025

This is the dataset for the ICDAR 2025 Competition on Comics Understanding in the Era of Foundational Models

The competition is hosted in the Robust Reading Competition website and the leaderboard is available here.

The dataset contains five subtask or skills:

Sequence Filling

Given a sequence of comic panels, a missing panel, and a set of option panels, the task is to pick the panel that best fits the sequence.

Character Coherence, Visual Closure, Text Closure

These skills require understanding the context sequence to then pick the best panel to continue the story, focusing on the characters, the visual elements, and the text:

• Character Coherence: Given a sequence of comic panels, pick the panel from the two options that best continues the story in a coherent with the characters. Both options are the same panel, but the text in the speech bubbles has been swapped.
• Visual Closure: Given a sequence of comic panels, pick the panel from the options that best continues the story in a coherent way with the visual elements.
• Text Closure: Given a sequence of comic panels, pick the panel from the options that best continues the story in a coherent way with the text. All options are the same panel, but with text in the speech retrieved from different panels.

Caption Relevance

Given a caption from the previous panel, select the panel that best continues the story.

Loading the Data

from datasets import load_dataset

skill = "sequence_filling" # "sequence_filling", "char_coherence", "visual_closure", "text_closure", "caption_relevance"
split = "val" # "val", "test", "train"
dataset = load_dataset("VLR-CVC/ComicsPAP", skill, split=split)

Map to single images

If your model can only process single images, you can render each sample as a single image:

from PIL import Image, ImageDraw, ImageFont
import numpy as np
from datasets import Features, Value, Image as ImageFeature

class SingleImagePickAPanel:
    def __init__(self, max_size=500, margin=10, label_space=20, font_path=None):
        if font_path is None:
            raise ValueError("Font path must be provided. Testing was done with 'Arial.ttf'")
        self.max_size = max_size
        self.margin = margin
        self.label_space = label_space
        # Add separate font sizes
        self.label_font_size = 20
        self.number_font_size = 24

        self.font_path = font_path

    def resize_image(self, img):
        """Resize image keeping aspect ratio if longest edge > max_size"""
        if max(img.size) > self.max_size:
            ratio = self.max_size / max(img.size)
            new_size = tuple(int(dim * ratio) for dim in img.size)
            return img.resize(new_size, Image.Resampling.LANCZOS)
        return img

    def create_mask_panel(self, width, height):
        """Create a question mark panel"""
        mask_panel = Image.new("RGB", (width, height), (200, 200, 200))
        draw = ImageDraw.Draw(mask_panel)
        font_size = int(height * 0.8)
        try:
            font = ImageFont.truetype(self.font_path, font_size)
        except:
            raise ValueError("Font file not found")
        
        text = "?"
        bbox = draw.textbbox((0, 0), text, font=font)
        text_x = (width - (bbox[2] - bbox[0])) // 2
        text_y = (height - (bbox[3] - bbox[1])) // 2
        draw.text((text_x, text_y), text, fill="black", font=font)
        return mask_panel

    def draw_number_on_panel(self, panel, number, font):
        """Draw number on the bottom of the panel with background"""
        draw = ImageDraw.Draw(panel)
        
        # Get text size
        bbox = draw.textbbox((0, 0), str(number), font=font)
        text_width = bbox[2] - bbox[0]
        text_height = bbox[3] - bbox[1]
        
        # Calculate position (bottom-right corner)
        padding = 2
        text_x = panel.size[0] - text_width - padding
        text_y = panel.size[1] - text_height - padding
        
        # Draw semi-transparent background
        bg_rect = [(text_x - padding, text_y - padding), 
                  (text_x + text_width + padding, text_y + text_height + padding)]
        draw.rectangle(bg_rect, fill=(255, 255, 255, 180))
        
        # Draw text
        draw.text((text_x, text_y), str(number), fill="black", font=font)
        return panel

    def map_to_single_image(self, examples):
        """Process a batch of examples from a HuggingFace dataset"""
        single_images = []
        
        for i in range(len(examples['sample_id'])):
            # Get context and options for current example
            context = examples['context'][i] if len(examples['context'][i]) > 0 else []
            options = examples['options'][i]
            
            # Resize all images
            context = [self.resize_image(img) for img in context]
            options = [self.resize_image(img) for img in options]
            
            # Calculate common panel size (use median size to avoid outliers)
            all_panels = context + options
            if len(all_panels) > 0:
                widths = [img.size[0] for img in all_panels]
                heights = [img.size[1] for img in all_panels]
                panel_width = int(np.median(widths))
                panel_height = int(np.median(heights))
                
                # Resize all panels to common size
                context = [img.resize((panel_width, panel_height)) for img in context]
                options = [img.resize((panel_width, panel_height)) for img in options]
                
                # Create mask panel for sequence filling tasks if needed
                if 'index' in examples and len(context) > 0:
                    mask_idx = examples['index'][i]
                    mask_panel = self.create_mask_panel(panel_width, panel_height)
                    context.insert(mask_idx, mask_panel)
                
                # Calculate canvas dimensions based on whether we have context
                if len(context) > 0:
                    context_row_width = panel_width * len(context) + self.margin * (len(context) - 1)
                    options_row_width = panel_width * len(options) + self.margin * (len(options) - 1)
                    canvas_width = max(context_row_width, options_row_width)
                    canvas_height = (panel_height * 2 + 
                                   self.label_space * 2)
                else:
                    # Only options row for caption_relevance
                    canvas_width = panel_width * len(options) + self.margin * (len(options) - 1)
                    canvas_height = (panel_height + 
                                   self.label_space)
                
                # Create canvas
                final_image = Image.new("RGB", (canvas_width, canvas_height), "white")
                draw = ImageDraw.Draw(final_image)
                
                try:
                    label_font = ImageFont.truetype(self.font_path, self.label_font_size)
                    number_font = ImageFont.truetype(self.font_path, self.number_font_size)
                except:
                    raise ValueError("Font file not found")
                
                current_y = 0
                
                # Add context section if it exists
                if len(context) > 0:
                    # Draw "Context" label
                    bbox = draw.textbbox((0, 0), "Context", font=label_font)
                    text_x = (canvas_width - (bbox[2] - bbox[0])) // 2
                    draw.text((text_x, current_y), "Context", fill="black", font=label_font)
                    current_y += self.label_space
                    
                    # Paste context panels
                    x_offset = (canvas_width - (panel_width * len(context) + 
                               self.margin * (len(context) - 1))) // 2
                    for panel in context:
                        final_image.paste(panel, (x_offset, current_y))
                        x_offset += panel_width + self.margin
                    current_y += panel_height
                
                # Add "Options" label
                bbox = draw.textbbox((0, 0), "Options", font=label_font)
                text_x = (canvas_width - (bbox[2] - bbox[0])) // 2
                draw.text((text_x, current_y), "Options", fill="black", font=label_font)
                current_y += self.label_space
                
                # Paste options with numbers on panels
                x_offset = (canvas_width - (panel_width * len(options) + 
                           self.margin * (len(options) - 1))) // 2
                for idx, panel in enumerate(options):
                    # Create a copy of the panel to draw on
                    panel_with_number = panel.copy()
                    if panel_with_number.mode != 'RGBA':
                        panel_with_number = panel_with_number.convert('RGBA')
                    
                    # Draw number on panel
                    panel_with_number = self.draw_number_on_panel(
                        panel_with_number, 
                        idx, 
                        number_font
                    )
                    
                    # Paste the panel with number
                    final_image.paste(panel_with_number, (x_offset, current_y), panel_with_number)
                    x_offset += panel_width + self.margin
                
                # Convert final_image to PIL Image format (instead of numpy array)
                single_images.append(final_image)
            
        # Prepare batch output
        examples['single_image'] = single_images
        
        return examples

from datasets import load_dataset

skill = "sequence_filling" # "sequence_filling", "char_coherence", "visual_closure", "text_closure", "caption_relevance"
split = "val" # "val", "test"
dataset = load_dataset("VLR-CVC/ComicsPAP", skill, split=split)

processor = SingleImagePickAPanel()
dataset = dataset.map(
        processor.map_to_single_image,
        batched=True,
        batch_size=32,
        remove_columns=['context', 'options']
    )
dataset.save_to_disk(f"ComicsPAP_{skill}_{split}_single_images")

Evaluation

The evaluation metric for all tasks is the accuracy of the model's predictions. The overall accuracy is calculated as the weighted average of the accuracy of each subtask, with the weights being the number of examples in each subtask.

To evaluate on the test set you must submit your predictions to the Robust Reading Competition website, as a json file with the following structure:

[
    { "sample_id" : "sample_id_0", "correct_panel_id" : 3},
    { "sample_id" : "sample_id_1", "correct_panel_id" : 1},
    { "sample_id" : "sample_id_2", "correct_panel_id" : 4},
    ...,
]

Where sample_id is the id of the sample, correct_panel_id is the prediction of your model as the index of the correct panel in the options.

Pseudocode for the evaluation on val set, adapt for your model:

skills = {
    "sequence_filling": {
        "num_examples": 262
    },
    "char_coherence": {
        "num_examples": 143
    },
    "visual_closure": {
        "num_examples": 300
    },
    "text_closure": {
        "num_examples": 259
    },
    "caption_relevance": {
        "num_examples": 262
    }
}

for skill in skills:
    dataset = load_dataset("VLR-CVC/ComicsPAP", skill, split="val")
    correct = 0
    total = 0
    for example in dataset:
        # Your model prediction
        prediction = model.generate(**example)
        prediction = post_process(prediction)
        if prediction == example["solution_index"]:
            correct += 1
        total += 1
    accuracy = correct / total
    print(f"Accuracy for {skill}: {accuracy}")

    assert total == skills[skill]["num_examples"]
    skills[skill]["accuracy"] = accuracy

# Calculate overall accuracy
total_examples = sum(skill["num_examples"] for skill in skills.values())
overall_accuracy = sum(skill["num_examples"] * skill["accuracy"] for skill in skills.values()) / total_examples
print(f"Overall accuracy: {overall_accuracy}")
    

Baselines

Find code for evaluation and training at: https://github.com/llabres/ComicsPAP

Baselines on the validation set using single images:

Model	Repo	Sequence Filling (%)	Character Coherence (%)	Visual Closure (%)	Text Closure (%)	Caption Relevance (%)	Total (%)
Random		20.22	50.00	14.41	25.00	25.00	24.30
Qwen2.5-VL-3B (Zero-Shot)	Qwen/Qwen2.5-VL-3B-Instruct	27.48	48.95	21.33	27.41	32.82	29.61
Qwen2.5-VL-7B (Zero-Shot)	Qwen/Qwen2.5-VL-7B-Instruct	30.53	54.55	22.00	37.45	40.84	34.91
Qwen2.5-VL-72B (Zero-Shot)	Qwen/Qwen2.5-VL-72B-Instruct	46.88	53.84	23.66	55.60	38.17	41.27
Qwen2.5-VL-3B (Lora Fine-Tuned)	coming soon	62.21	93.01	42.33	63.71	35.49	55.55
Qwen2.5-VL-7B (Lora Fine-Tuned)	coming soon	69.08	93.01	42.00	74.90	49.62	62.31

Citation

coming soon