QMUL Attributes for Re-identification

We present here our attributes ground-truth labels, as used in the below publications. Attributes are a compelling representation for re-identification; they're low-dimensional, often easy to classify, and lastly they are quite compatible with other standard tools such as distance-metric learning and transfer.

Details

We present here our attribute ground-truths from our BMVC2012 and recent Springer publications:

"Attributes-based Re-identification"

R. Layne, T.M. Hospedales and S. Gong

In Gong, Cristani, Yan, Loy (Eds.), Person Re-Identification, Springer, December 2013

Read our REID chapter as PDF Download annotations, data, and example script
Full text at Springer.com

This ontology comprises of 21 different attributes, selected for their classifiability and discriminative qualities, as well as human domain experts.

Specs

  • 21 attributes, in order of prevalence in VIPeR: (darkhair, hassatchel, darkbottoms, midhair, greenshirt, darkshirt, lightshirt, bald, jeans, lightpants, uniquepants, nocoat, redshirt, hasbaginhand, barelegs, skirt, ismale, hasbackpack, blueshirt, patterned, shorts).
  • Labels provided for VIPeR, GRID, PRID datasets (provided in .mat format)
  • Example script provided, illustrating best-case re-identification scores vs. standard L2 nearest-neighbour re-id.

"Person Re-Identification by Attributes"

R. Layne, T.M. Hospedales and S. Gong

British Machine Vision Conference, Surrey, England, 2012

PDF Download Annotations

This file comprises our original 15 attributes, selected solely on the operating principles of human experts.

Specs

  • 15 attributes, 12 of which are appearance-based, the remaining 3 are soft-biometrics: (shorts, skirt, sandals, backpack, jeans, logo, v-neck, open-outerwear, stripes, sunglasses, headphones, long-hair, short-hair, gender, carrying- object).
  • Labels provided for VIPeR, and PRID datasets (provided in .mat format)