The PigDetect dataset was designed to support the benchmarking of pig detection algorithms from images as introduced in the "A Large-Scale Longitudinal Dataset for Pig Tracking and Re-Identification" article. To ensure both diversity and representation, two data allocation strategies were applied. The first followed the standard practice by randomly sampling frames across the dataset. The second method targeted high-density scenarios by ranking frames according to the mean Intersection over Union (IoU) of all objects within each frame, and removing subsequent neighbouring frames (N = 5) to avoid over‑representation of similar viewpoints. From the selected pool of images, the top 200 frames were retained for subsequent random sampling. For both allocation strategies, the same imaging days were used to define the training/validation/test splits. Fourteen imaging days (representing different groups, ages, and pen sizes) were assigned to the test subset (N = 280 frames), while the remaining 202 days formed the training and validation pool (N = 4,040 frames). For each imaging day and group, 20 frames were randomly selected from the curated subset. The final training/validation data was partitioned using a 90:10 split, while the testing subset was generated using separate held-out days, representing approx. 7% of the combined training and validation sets.