kaityc06/NABirds_masked

NABirds Masked

Segmentation masks for the NABirds dataset (training split), stored as RLE-encoded masks in a single Parquet file.

File

File	Description
masks.parquet	20,694 rows — one per image — with RLE mask, score, Qwen quality verdict, species label, and image_id

Schema

Column	Type	Description
image_id	str	Sequential identifier for the image (e.g. NABirds_000000)
dataset	str	Always NABirds
label	str	Raw label string (e.g. species='Eared Grebe (Nonbreeding/juvenile)')
text_prompt	str	Text prompt used to generate the mask (always bird)
mask_rle_counts	str	RLE-encoded mask (COCO format)
mask_rle_height	int	Height used for RLE decoding
mask_rle_width	int	Width used for RLE decoding
mask_score	float	Mask confidence score
mask_area_frac	float	Fraction of image area covered by the mask
image_width	int	Original image width
image_height	int	Original image height
qwen_verdict	bool	Whether the Qwen3.5 quality checker approved the mask
qwen_analysis	str	Qwen3.5 analysis text explaining the verdict
species	str	Species common name (e.g. Eared Grebe (Nonbreeding/juvenile))
bbox	str	Bounding box [x0, y0, x1, y1] normalised to [0, 1], or null if unavailable

Loading the masks

import pandas as pd

df = pd.read_parquet("hf://datasets/kaityc06/NABirds_masked/masks.parquet")

Retrieving the original image by species

The species column matches the class names in the NABirds classes.txt file. Build a lookup from species name to image paths once, then open any image by its species:

import os
import pandas as pd
from PIL import Image

# Path to your local NABirds download (must contain images.txt, classes.txt, etc.)
NABIRDS_DIR = "/path/to/nabirds"

def _read_table(filename):
    rows = {}
    with open(os.path.join(NABIRDS_DIR, filename)) as f:
        for line in f:
            parts = line.strip().split(" ", 1)
            rows[parts[0]] = parts[1]
    return rows

images      = _read_table("images.txt")
labels      = _read_table("image_class_labels.txt")
class_names = _read_table("classes.txt")

# Build species name → list of absolute image paths
species_to_paths = {}
for img_id, rel_path in images.items():
    species = class_names.get(labels[img_id], "unknown")
    abs_path = os.path.join(NABIRDS_DIR, "images", rel_path)
    species_to_paths.setdefault(species, []).append(abs_path)

# Look up any species from the parquet
species = "Eared Grebe (Nonbreeding/juvenile)"
img = Image.open(species_to_paths[species][0])

Decoding a mask

import numpy as np
from pycocotools import mask as mask_utils

row = df.iloc[0]

rle = {
    "counts": row["mask_rle_counts"],
    "size": [row["mask_rle_height"], row["mask_rle_width"]],
}
binary_mask = mask_utils.decode(rle)  # numpy array, shape (H, W), dtype uint8