metadata
dataset_info:
features:
- name: image
dtype: image
- name: image_id
dtype: int64
- name: file_name
dtype: string
- name: width
dtype: int64
- name: height
dtype: int64
- name: camera_height_img
dtype: int64
- name: objects
struct:
- name: annotation_id
list: int64
- name: area
list: float64
- name: bbox
list:
list: float64
- name: camera_height_ann
list: int64
- name: category_id
list: int64
- name: difficult
list: int64
- name: ignore
list: int64
- name: person_location
list:
list: float64
- name: ritbox
list:
list: float64
- name: rotated_box
list:
list: float64
- name: segmentation
list:
list: float64
- name: world_location
list:
list: float64
splits:
- name: train
num_bytes: 2908124774.58
num_examples: 29569
- name: validation
num_bytes: 557308312.6
num_examples: 4600
- name: test
num_bytes: 1173421936.14
num_examples: 8773
download_size: 4870084026
dataset_size: 4638855023.32
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
from datasets import load_dataset
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import numpy as np
# 1. Load the dataset
# Note: Since this is a private repo, ensure you have run `huggingface-cli login`
repo_id = "bdanko/loaf_resolution_512"
print(f"Downloading {repo_id}...")
dataset = load_dataset(repo_id, split="train")
# 2. Grab the first example
example = dataset[0]
# Hugging Face automatically decodes the Parquet bytes into a PIL Image
img = example["image"]
objects = example["objects"]
# Print image-level metadata
print("\n--- Image Metadata ---")
print(f"File Name: {example['file_name']}")
print(f"Image ID: {example['image_id']}")
print(f"Dimensions: {example['width']}x{example['height']}")
print(f"Camera Height (Image): {example['camera_height_img']}")
# 3. Set up the matplotlib plot
fig, ax = plt.subplots(1, figsize=(10, 8))
ax.imshow(img)
ax.axis("off")
ax.set_title(f"Sample: {example['file_name']}", fontsize=14)
print("\n--- Object Metadata ---")
# 4. Iterate through all objects in this image
# We use zip to unpack all the parallel lists stored inside the 'objects' dictionary
num_objects = len(objects["annotation_id"])
for i in range(num_objects):
ann_id = objects["annotation_id"][i]
cat_id = objects["category_id"][i]
bbox = objects["bbox"][i]
seg = objects["segmentation"][i]
# Custom / 3D annotations
ritbox = objects["ritbox"][i]
rot_box = objects["rotated_box"][i]
world_loc = objects["world_location"][i]
person_loc = objects["person_location"][i]
cam_height_ann = objects["camera_height_ann"][i]
print(f"\nObject {i+1} (Ann ID: {ann_id} | Category: {cat_id})")
print(f" - 2D BBox: {bbox}")
print(f" - Ritbox: {ritbox}")
print(f" - Rotated Box: {rot_box}")
print(f" - World Location: {world_loc}")
print(f" - Person Location: {person_loc}")
print(f" - Camera Height: {cam_height_ann}")
# --- Plotting the 2D Bounding Box ---
# COCO bbox format is [top_left_x, top_left_y, width, height]
if len(bbox) == 4:
x, y, w, h = bbox
rect = patches.Rectangle(
(x, y), w, h,
linewidth=2, edgecolor='red', facecolor='none', label=f"Cat {cat_id}"
)
ax.add_patch(rect)
ax.text(x, y - 5, f"Cat {cat_id}", color='red', fontsize=10, weight='bold')
# --- Plotting the Segmentation Polygon ---
# COCO segmentation is a flat list: [x1, y1, x2, y2, ...]
if len(seg) > 0:
# Reshape flat list into a list of (x, y) coordinate pairs
poly_coords = np.array(seg).reshape(-1, 2)
polygon = patches.Polygon(
poly_coords,
linewidth=2, edgecolor='cyan', facecolor='cyan', alpha=0.3
)
ax.add_patch(polygon)
plt.tight_layout()
plt.show()
