afmck/peanuts-opt-6.7b

Peanut Comic Strip Dataset (Snoopy & Co.)

This is a dataset Peanuts comic strips from 1950/10/02 to 2000/02/13. There are 77,457 panels extracted from 17,816 comic strips. The dataset size is approximately 4.4G.

Each row in the dataset contains the following fields:

• image: PIL.Image containing the extracted panel.
• panel_name: unique identifier for the row.
• characters: tuple[str, ...] of characters included in the comic strip the panel is part of.
• themes: tuple[str, ...] of theme in the comic strip the panel is part of.
• color: str indicating whether the panel is grayscale or in color.
• caption: BLIP-2_OPT_6.7B generated caption from the panel.
• year: int storing the year the specific panel was released.

OPT-6.7B has a non-commercial use license and so this dataset cannot be used for commercial projects. If you need a dataset for commercial use please see this similar dataset that uses Flan-T5-XL, which allows for commercial use.

Character and theme information was extracted from Peanuts Wiki (Fandom) using Beautiful Soup. Images were extracted from Peanuts Search.

Only strips with the following characters were extracted:

- "Charlie Brown"
- "Sally Brown"
- "Joe Cool" # Snoopy alter-ego
- "Franklin"
- "Violet Gray"
- "Eudora"
- "Frieda"
- "Marcie"
- "Peppermint Patty"
- "Patty"
- "Pig-Pen"
- "Linus van Pelt"
- "Lucy van Pelt"
- "Rerun van Pelt"
- "Schroeder"
- "Snoopy"
- "Shermy"
- "Spike"
- "Woodstock"
- "the World War I Flying Ace" # Snoopy alter-ego

Extraction Details

Panel detection and extraction was done using the following codeblock:

def check_contour(cnt):
    area = cv2.contourArea(cnt)
    if area < 600:
        return False

    _, _, w, h = cv2.boundingRect(cnt)
    if w / h < 1 / 2: return False
    if w / h > 2 / 1: return False

    return True

def get_panels_from_image(path):
    panels = []
    original_img = cv2.imread(path)
    gray = cv2.cvtColor(original_img, cv2.COLOR_BGR2GRAY)
    blur = cv2.GaussianBlur(gray, (5,5), 0)
    thresh = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1]

    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3,3))
    opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations=1)
    invert = 255 - opening

    cnts, _ = cv2.findContours(invert, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)

    idx = 0 
    for cnt in cnts:
        if not check_contour(cnt): continue
        idx += 1
        x,y,w,h = cv2.boundingRect(cnt)
        roi = original_img[y:y+h,x:x+w]
        panels.append(roi)

    return panels

check_contour will reject panels with area < 600 or with aspect ratios larger than 2 or smaller than 0.5.

Grayscale detection was done using the following codeblock:

def is_grayscale(panel):
    LAB_THRESHOLD = 10.
    img = cv2.cvtColor(panel, cv2.COLOR_RGB2LAB)
    _, ea, eb = cv2.split(img)
    de = abs(ea - eb)
    mean_e = np.mean(de)
    return mean_e < LAB_THRESHOLD

Captioning was done using the standard BLIP-2 pipeline shown in the Huggingface docs using beam search over 10 beams and a repetition penalty of 2.0. Raw captions are extracted and no postprocessing is applied. You may wish to normalise captions (such as replacing "cartoon" with "peanuts cartoon") or incorporate extra metadata into prompts.