Preprocessed CR-39 detector particle track image dataset with annotations

First, manual labeling inherently involves subjective judgment, particularly in defining precise object boundaries. Second, for complete (non-occluded) particles, annotators were required to infer the full particle shape when portions of the particle boundary were ambiguous or partially obscured, introducing additional uncertainty in the ground-truth box annotations.Two categories of particle data are included: SET1 consists of complete (non-occluded) particles, and SET2 consists of partially occluded particles.

A subset of SET1 was used for model training, while another subset served as the in-distribution test set and validation set for hyperparameter tuning. SET2 was used entirely as the out-of-distribution prediction set, as these occluded samples lack strict ground truth.The training set comprises 95 images at etching time T = 1 h and 269 images at T = 3 h. All images were rotated from 0° to 330° in 30° increments, yielding 4368 standardized database entries.

Each entry includes the standardized 128 × 128 pixel image (resized via bilinear interpolation) and the original image dimensions (a 1 × 2 tensor). Three test sets of increasing difficulty are constructed: Test Set 1 (In-Distribution, sample size: 4368) consists of database images with real occlusion patterns randomly sampled from actual particle images; Test Set 2 (Out-of-Distribution, also used as validation set for hyperparameter selection, sample size: 840 for T = 1 h and 2244 for T = 3 h) consists of complete particle images not present in the database with the same occlusion patterns applied; Test Set 3 (Real Incomplete, sample size: 1272, with 162 for T = 1 h and 1110 for T = 3 h) contains genuinely incomplete particle tracks extracted from real environments without corresponding complete ground truth.The uploaded data files are in .pt format (PyTorch tensor format), containing the preprocessed 128 × 128 images and associated metadata.

Each .pt file contains a dictionary with keys: 'image' (the 128×128 tensor), 'original_dims' (the 1×2 tensor of original dimensions), and 'label' (the associated physical parameters,here is the etch time T in txt). The data are organized according to the train/validation/test splits described above, with separate directories for each set. No data are missing; the primary errors originate from manual annotation uncertainty and the inference of complete particle boundaries as described above, both of which are within acceptable tolerance for this task.