K-reciprocal Encoding

Introduction

• This is my python implementation of a CVPR-2017 paper Re-ranking Person Re-identification with k-reciprocal Encoding
• The authors have released their codes here in github too.

How to run?

• python3 main.py
• Check main.py and replace gallery_X, gallery_Y, query_X, query_X with any data you want to try on.

Performance

• The code is tested on a small portion of LFW dataset.
• mAP = 0.1553 using Euclidean Distance w Re-ranking
• mAP = 0.1025 using Euclidean Distance w/o Re-ranking

Problems to resolve

• There are some parts in the original code which I do not understand.
• Those parts are not implemented in my code. Hence my implementation is a little but faster, but the performace is quite worse.
• mAP = 0.2182 using Euclidean Distance w the original code

Problem 1

original_dist = np.concatenate(
    [np.concatenate([q_q_dist, q_g_dist], axis=1),
    np.concatenate([q_g_dist.T, g_g_dist], axis=1)],
axis=0)

• The probe and gallery are mixed here.
• One probe can be another probe's k-nearest neighbors.
• One probe can be another probe's k-reciprocal nearest neighbors.
• The overall computation of N, R and V are not independent within probes.

Problem 2

original_dist = np.power(original_dist, 2).astype(np.float32)
original_dist = np.transpose(1. * original_dist/np.max(original_dist,axis = 0))

• Why?

Problem 3

weight = np.exp(-original_dist[i,k_reciprocal_expansion_index])

• In the paper, the value is exp(-d) instead of whatever this is.