Dataset/README.md

Dataset

Raw photos live in items/; the paired evaluation set (same/, different/, pairs.csv) is generated from them by python scripts/build_dataset.py and is not committed.

Raw photos: items/

items/<item>/front.jpeg, back.jpeg, side.jpeg          photo set 1
items/<item>/set2/front.jpeg, back.jpeg, side.jpeg     photo set 2, taken separately (only some items)

Item	What it is	Photo sets
speaker_A	clean blue Portronics speaker with a batch label on the base	1
speaker_B	same model, worn black unit, label gone	1
shaker_A / shaker_B	two different shaker bottles	1
bottle_A / bottle_B	same steel bottle model; B has a strap and small dents	1
headphones_blue	blue Sony over-ear headphones	2
headphones_black	black, worn Sony headphones of the same design	2 (set 2 has a new front photo only; back and side reuse set 1)

Angles: front = the main view (earcups facing the camera, product upright), back = the reverse / underside view, side = the side profile (for headphones: the slider arm with the SONY logo).

Pairs: same/ and different/

<same|different>/delivery_images/<front|back|side>/<pair_id>.jpeg
<same|different>/return_images/<front|back|side>/<pair_id>.jpeg

Pair IDs are numbered separately in each folder. same/.../25.jpeg and different/.../25.jpeg are unrelated pairs. Within one pair, the delivery and return files share the same number. pairs.csv lists every pair: label, pair_id, delivery_item, return_item, pair_type, delivery_set, return_set.

Folder	IDs	Contents
same	1–18	the first 6 items, 3 pairs each: return photo = the delivery photo, digitally re-captured (rotation, zoom, exposure, white balance, blur, JPEG)
same	19–24	the headphones, the same way
same	25–28	real re-photographs: delivery from one photo set, return from the other (25 & 27: set 1 → set 2; 26 & 28: set 2 → set 1). They look different on purpose
different	1–30	every ordered pair of the first 6 items
different	31–56	every new ordered pair involving the headphones

Return images in different/ get the same simulated re-capture as in same/, so the two classes differ only by the item. Seeds are fixed: rebuilding produces byte-identical files, and adding items never changes existing pairs.