Real-Time Pokémon Card Detection from Tournament Footage - Object Detection and Dataset Augmentation Methods

Sep 11, 2026

Blurred, fast-moving Pokémon tournament footage makes card reading difficult for many viewers, motivating a computer-vision approach for real-time recognition. The work targets accurate detection of multiple Pokémon card classes by combining established data augmentation with new dataset-generation ideas. Techniques include random rotations and copy-paste augmentation, plus mosaic backgrounds and polygon clipping for training realism. The paper also discusses practical limitations of the detection pipeline under real tournament conditions.

LanguageEnglish

CategoryPresentations

FormatPDF

Pages9

Reading Time3 min

AccessFree to download

Introduction

Related Works

Dataset

Methods

Real-Time Pokémon Card Detection from Tournament Footage Edwin Pua Department of Computer Science Stanford University [puaedwin@stanford.edu](puaedwin@stanford.edu) Abstract Match footage for Pokémon tournaments is blurry and hard to follow for average viewers unfamiliar with the cards. This paper outlines techniques dataset augmentation to improve an object detection model’s ability to detect classes of Pokémon cards in real time, as well as their limitations. This includes established techniques such as random rotations or copy-paste augmentation, as well as novel techniques such as mosaic backgrounds or polygon clipping. 1 Introduction The most popular use of Pokémon Trading Cards is for collecting, but they have another use-battling. Pokémon Trading Cards are used to play a tabletop game that simulates Pokémon battles. However, there are two primary barriers that prevent people from getting into the sport. One is the set of twenty minute rules required to learn how to play, which is a surprisingly high activation cost. The other is how difficult it is to be excited as a spectator of the Pokémon Card Game. Figure 1: Example of tournament footage of the Pokémon Trading Card Game (TCG) . For instance, consider Figure 1 . In order to capture the whole playing field, the camera needs to be a certain distance above the match. However, because the Pokémon and text on the cards are illegible from this distance, the only way to follow what is happening is if you know what all the cards doin advance. Most normal spectators will not understand what is happening, which might turn away potential Pokémon trainers. To alleviate this, the goal is to make a computer vision model that can detect which cards are onscreen. When a viewer hovers over a particular card, the plan is to show a zoomed-in, high quality version of the card so an online spectator can easily read what a card does. To be more specific, given a URL or MP4 recording of Pokémon TCG tournament footage, is it possible to generate real-time, frame-by-frame segmentation masks of each Pokémon card as output? Stanford CS224N Natural Language Processing with Deep Learning That way, when a user uses the mouse to hover over a card, and the cursor coordinates coincide with one of the coordinates of a segmentation mask, we can use the detected card class of that segmentation mask to retrieve the correct zoomed in image of the corresponding card. 2 Related Works In the past, there have been similar works on card detection. For instance, Snyder developed a poker card detector without using a neural network, via corner detection and template matching. [1] However, these heuristics are not robust, since they assume close-up, clean images and no perspective distortion of the cards. Another poker card detector was made by Chen et al. using a "sandglass block", or encoder-bottleneck-decoder network architecture, which works even when each card takes up only 0.7% of the screen. However, these poker cards are not occluded by any objects like cards in the Pokémon card game are (e.g. dice, tokens, etc.), and there are only 52 classes of poker cards, compared to the thousands of classes of Pokémon cards. [2] Generalizing to overall object detection, perhaps one of the most seminal papers is the YOLO paper by Redmon et al. [3] This model specializes in single-shot, real-time object detection, which is perfect for my use case. Given that I can easily finetune a YOLO model on my specific task, I feel confident in choosing YOLO as my weapon of choice. Given that there exists only one of each Pokémon card, data augmentation will be very important for generating datasets. For this reason, I plan to implement many established augmentation techniques, such as random, noise, flip, rotation, etc. posited by Yang et al. [4] Another technique is copypaste augmentation and scale jittering, as presented by Ghiasi et al. [5] I could also implement the techniques proposed

FAQs

What problem does the paper address in Pokémon tournament viewing?

Tournament footage is often blurry and the cards are hard to read for average viewers, reducing the spectator experience and making card understanding difficult without knowing the card effects in advance.

How does the proposed system help viewers read cards?

A model detects which cards are on screen and produces segmentation masks so that hovering with a cursor can trigger a zoomed-in, high-quality card view for easier reading.

What data and augmentation strategies are emphasized?

The paper constructs training data using web-scraped tournament-legal cards and builds validation data from real tournament screenshots, then improves robustness with augmentations such as random rotations, copy-paste, mosaic backgrounds, and polygon clipping.

Sep 11, 2026

Uploaded by Clementine