File size: 7,125 Bytes
c679d56 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 | """
Module for dataset and dataloaders of UCSD dataset.
"""
import re
from pathlib import Path
from typing import List, Tuple, Optional
import numpy as np
from PIL import Image
import torch
from torch.utils.data import Dataset
from src.data.video_transforms import transform
class UCSDDataset(Dataset):
"""
UCSD Anomaly Detection Dataset.
Train: only normal clips.
Test: clips with frame-level ground truth annotations.
Args:
root: Dataset root path (containing UCSDped1/, UCSDped2/)
subset: 'Ped1' or 'Ped2'
split: 'train' or 'test'
window_size: Number of frames per sample (sliding window)
stride: Stride between windows
transform: Optional transform applied to each frame
"""
def __init__(
self,
root: str,
subset: str = "Ped2",
split: str = "train",
window_size: int = 16,
stride: int = 8,
mode: str = "reconstruction",
transform: Optional[callable] = None,
clip_indices: Optional[List[int]] = None
):
super().__init__()
self.root = Path(root)
self.subset = subset.lower()
self.split = split
self.window_size = window_size
self.stride = stride
self.mode = mode
self.transform = transform
self.clip_indices = clip_indices
# Subset check
assert subset in ("ped1", "ped2"), f"subset must be ped1 or ped2, got {subset}"
# Read the subset and store the clip directories
self.subset_split = self.root / f"UCSD{self.subset}" / f"{split.title()}"
# Sanity check to ensure the files and clip directories exist
if not self.subset_split.exists():
raise FileNotFoundError(f"Dataset path not found: {self.subset_split}")
self.clip_dirs = sorted([
d for d in self.subset_split.iterdir()
if d.is_dir() and not d.name.endswith("_gt")
])
# Filter out the clips
if self.clip_indices is not None:
self.clip_dirs = [self.clip_dirs[i] for i in self.clip_indices]
if len(self.clip_dirs) == 0:
raise RuntimeError(f"No clip directories found in {self.subset_split}")
# Collect the clip paths
self.clips = []
for clip_dir in self.clip_dirs: # clip_dir = Path("Train001")
frame_paths = sorted(clip_dir.glob("*.tif")) # liste of frame paths
frames = np.stack([np.array(Image.open(p)) for p in frame_paths])
self.clips.append(frames)
# Create labels based on split
if self.split == "test":
m_file = self.subset_split / f"UCSD{subset}.m" # path case dikkat
content = m_file.read_text()
matches = re.findall(r"\[(\d+):(\d+)\]", content)
self.labels = []
for clip_idx, (start_str, end_str) in enumerate(matches):
start, end = int(start_str), int(end_str)
n_frames = len(self.clips[clip_idx]) # clip's frame length
label = np.zeros(n_frames, dtype=np.int64)
label[start-1:end] = 1 # 1-indexed -> 0-indexed slice
self.labels.append(label)
else:
self.labels = None # train, no label
# Collect the window indexes
self.windows = [] # list of (clip_idx, start_frame)
for clip_idx, frames in enumerate(self.clips):
n_frames = len(frames)
for start in range(0, n_frames - window_size + 1, stride):
self.windows.append((clip_idx, start))
def __len__(self) -> int:
return len(self.windows)
def __getitem__(self, idx: int) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Returns:
frames: (T, C, H, W) tensor
label: (T,) tensor of 0/1 (train: all zeros, test: from gt)
"""
# Read frames and label
clip_idx, start_frame = self.windows[idx]
# Take the frames within frame range
window_frames = self.clips[clip_idx][start_frame : start_frame + self.window_size] # shape: (T, H, W) uint8
# Check labels based on split
if self.split == "test":
labels_np = self.labels[clip_idx][start_frame : start_frame + self.window_size]
labels = torch.from_numpy(labels_np) # int64 tensor
else:
labels = torch.zeros(self.window_size, dtype=torch.long)
# Convert window array to tensor and reshape it
window_tensor = torch.from_numpy(window_frames).float() / 255.0
window_tensor = window_tensor.unsqueeze(1) # (T, H, W) -> (T, 1, H, W)
# Check for transforms
if self.transform is not None:
window_tensor = self.transform(window_tensor)
if self.mode == "prediction":
input_frames = window_tensor[:-1] # (15, 1, H, W) — first 15 window
target_frame = window_tensor[-1] # (1, H, W) — last frame, target
return input_frames, target_frame
else:
return window_tensor, labels
if __name__ == "__main__":
# Run sanity check
train_clips = [0,1,2,3,4,5,6,7,8,9,10,11,12] # 13 clip
val_clips = [13,14,15] # 3 clip
# Train
ds_train = UCSDDataset(root="data/ucsd/raw", subset="ped2", clip_indices=train_clips, transform=transform, split="train")
print(f"Train: {len(ds_train.clips)} clips, {len(ds_train)} windows")
print(f"First clip shape: {ds_train.clips[0].shape}")
# Validation
ds_val = UCSDDataset(root="data/ucsd/raw", subset="ped2", clip_indices=val_clips, transform=transform, split="train")
print(f"Val: {len(ds_val.clips)} clips, {len(ds_val)} windows")
print(f"Val labels: {ds_val.labels}") # Should be None
# Test
ds_test = UCSDDataset(root="data/ucsd/raw", subset="ped2", split="test", transform=transform)
print(f"Test: {len(ds_test.clips)} clips, {len(ds_test)} windows")
print(f"First label sum: {ds_test.labels[0].sum()}/{len(ds_test.labels[0])}")
# Test getitem
sample, label = ds_train[0]
print(f"\nSample 0 (train):")
print(f" Sample shape: {sample.shape}, dtype: {sample.dtype}")
print(f" Sample range: [{sample.min():.3f}, {sample.max():.3f}]")
print(f" Label shape: {label.shape}, sum: {label.sum()}")
sample, label = ds_val[0]
print(f"\nSample 0 (test):")
print(f" Sample shape: {sample.shape}")
print(f" Label shape: {label.shape}, sum: {label.sum()}")
# Random middle sample
sample, label = ds_train[len(ds_train) // 2]
print(f"\nMiddle train sample shape: {sample.shape}")
# Transform check
print(sample.shape) # torch.Size([16, 1, 128, 128])
# Prediction
ds = UCSDDataset(root="data/ucsd/raw", subset="ped2", split="train",
clip_indices=list(range(13)), transform=transform, mode="prediction")
inp, tgt = ds[0]
print(f"input: {inp.shape}") # expected (15, 1, 128, 128)
print(f"target: {tgt.shape}") # expected (1, 128, 128) |