duclvQ/tcg-card-detector

TCG card corner detector

Finds a trading card in a photo and returns its four corners — top-left, top-right, bottom-right, bottom-left — so you can flatten the card to a front-facing view with one cv2.getPerspectiveTransform.

The corner order is fixed by keypoint index, so the output is ready to use: no sorting step, no ambiguity about which point is which.

Usage

pip install ultralytics huggingface_hub opencv-python

import cv2, numpy as np
from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download("duclvQ/tcg-card-detector", "card_pose_yolov8n.pt")
model = YOLO(weights)
img = cv2.imread("photo.jpg")

res = model.predict(img, imgsz=480, conf=0.25)[0]
if len(res.boxes):
    best = int(res.boxes.conf.argmax())
    quad = res.keypoints.xy.cpu().numpy()[best]     # (4, 2) TL, TR, BR, BL

    dst = np.float32([[0, 0], [200, 0], [200, 280], [0, 280]])
    M = cv2.getPerspectiveTransform(quad.astype(np.float32), dst)
    card = cv2.warpPerspective(img, M, (200, 280))  # flat 200x280 card

quad is [[x_tl, y_tl], [x_tr, y_tr], [x_br, y_br], [x_bl, y_bl]] in pixels of the original image.

For more than one card in a photo, iterate instead of taking the best box:

for quad in res.keypoints.xy.cpu().numpy():
    ...

Nano-sized: 6.4 MB, ~36 ms per image at imgsz=480 on a desktop GPU.

card_pose_480.onnx (ONNX Runtime) and card_pose_480.mlmodel (CoreML, iOS) are the same model exported at the same input size. The CoreML file has not been verified on Apple hardware — test it before shipping.

On real photographs

Every image on this page is a real photograph — a physical card, real lighting, real background. None are rendered or composited.

On a set of 186 photographs of Pokémon cards, 183 produced a usable quadrilateral. Cards from other games are found too, as above — the model reads the card's shape, not any one game's artwork.

A card that stands clear of its surroundings is what this model is good at. It is much weaker once cards pile up; see Limitations.

Limitations

Crowded scenes are the weak spot. Cards that are clearly separated are found confidently. Fanned hands, stacks, binder pages and heavy overlap usually return nothing at all.

A card with no corner in the frame cannot be found. An extreme macro crop leaves nothing to locate.

Plain, borderless card-like rectangles are less reliable. Real cards from every game tested are found; a card-shaped object with no border at all is much weaker.

Heavy blur plus occlusion — a card behind a hand, badly out of focus — is also missed.

Photo credits

The example photographs are from Wikimedia Commons and keep their own licences, which are separate from the model's. Attribution as those licences require:

photo	by	licence
Digimon_Omnimon_X_Anti_body_Card.jpg	slgckgc	CC BY 2.0
Japanese_magic_the_gathering_cards_block.jpg	Like_the_Grand_Canyon	CC BY-NC 2.0
Digimon_Omnimon_X_Anti_body_Card.jpg	slgckgc	CC BY 2.0
Brian_Kibler_USA_vs_Daniel_Antoniou_Cyprus_jpg.jpg	EuropeanTalent	CC BY 4.0
Bruce_Spraggins_a_logistics_planner_with_Multinational.jpg	1st Lt. Casey Staheli	Public domain
Daniel_Antoniou_jpg.jpg	EuropeanTalent	CC BY 4.0

Licence

AGPL-3.0, inherited from Ultralytics YOLOv8. Commercial use on other terms requires a licence from Ultralytics.

Trained on images of Pokémon cards, which are copyright The Pokémon Company / Nintendo. No card artwork is distributed here — this repository contains only detector weights, which locate a card-shaped quadrilateral and do not reproduce card images.