Note

Torchreid is a library for deep-learning person re-identification, written in PyTorch https://pytorch.org/ and developed for our ICCV'19 project, Omni-Scale Feature Learning for Person Re-Identification https://arxiv.org/abs/1905.00953.

1. Citation Repository

https://github.com/KaiyangZhou/deep-person-reid

2. ONNX Export

https://github.com/PINTO0309/deep-person-reid

3. Code snippet for calculating Cosine similarity (COS similarity) from feature vectors

Cosine similarity is calculated by dividing the inner product of two vectors by the product of their norms. However, since the vectors here are already normalized, simply computing the inner product results in the cosine similarity.

import torch
import torch.nn.functional as F

# Obtain feature vectors from images
with torch.no_grad():
    f1 = model(image1)  # Feature vector of image1
    f2 = model(image2)  # Feature vector of image2
    # Normalize and convert each vector to the unit norm (length is 1)
    A_feat = F.normalize(f1, dim=1).cpu()
    B_feat = F.normalize(f2, dim=1).cpu()
simlarity = A_feat.matmul(B_feat.transpose(1, 0)) # inner product of feature vectors
print("\033[1;31m The similarity is {}\033[".format(simlarity[0, 0]))

4. Similarity validation

	image.1	image.2
30 vs 31⬇️
30 vs 1⬇️
31 vs 2⬇️
1 vs 2⏫
1 vs 3⏫
1 vs 4⏫

python validation.py

Model	30 vs 31 ⬇️	30 vs 1 ⬇️	31 vs 2 ⬇️	1 vs 2 ⏫	1 vs 3 ⏫	1 vs 4 ⏫
1. mlfn
9cb5a267	0.521	0.558	0.609	0.609	0.725	0.740
2. mobilenetv2
1dot0_duke	0.496	0.542	0.501	0.654	0.852	0.773
1dot0_market	0.402	0.556	0.469	0.781	0.886	0.882
1dot0_msmt	0.522	0.412	0.471	0.678	0.624	0.621
1dot4_duke	0.518	0.633	0.552	0.729	0.853	0.779
1dot4_market	0.409	0.574	0.516	0.717	0.857	0.839
1dot4_msmt	0.503	0.430	0.425	0.629	0.651	0.714
1	0.430	0.433	0.423	0.427	0.428	0.429
3. osnet
ain_d_m_c	0.438	0.437	0.418	0.610	0.692	0.620
ain_ms_d_c	0.425	0.387	0.422	0.641	0.645	0.692
ain_ms_d_m	0.436	0.479	0.406	0.585	0.650	0.670
ain_ms_m_c	0.460	0.393	0.381	0.547	0.706	0.663
ain_x0_25_imagenet	0.546	0.362	0.448	0.554	0.703	0.669
ain_x0_5_imagenet	0.602	0.508	0.525	0.589	0.637	0.670
ain_x0_75_imagenet	0.522	0.529	0.477	0.643	0.686	0.716
ain_x1_0_dukemtmcreid_256x128_amsgrad_ep90_lr0	0.509	0.488	0.378	0.506	0.685	0.628
ain_x1_0_imagenet	0.504	0.500	0.491	0.579	0.750	0.720
ain_x1_0_market1501_256x128_amsgrad_ep100_lr0	0.426	0.540	0.461	0.582	0.825	0.785
ain_x1_0_msmt17_256x128_amsgrad_ep50_lr0	0.444	0.353	0.349	0.514	0.631	0.517
d_m_c	0.400	0.480	0.446	0.492	0.668	0.628
ibn_d_m_c	0.376	0.488	0.432	0.512	0.639	0.626
ibn_ms_d_c	0.440	0.428	0.373	0.642	0.678	0.633
ibn_ms_d_m	0.464	0.454	0.462	0.630	0.690	0.686
ibn_ms_m_c	0.439	0.432	0.467	0.575	0.701	0.616
ibn_x1_0_duke_256x128_amsgrad_ep150_stp60_lr0	0.423	0.425	0.440	0.506	0.703	0.639
ibn_x1_0_imagenet	0.549	0.495	0.552	0.536	0.761	0.720
ibn_x1_0_market1501_256x128_amsgrad_ep150_stp60_lr0	0.361	0.460	0.535	0.713	0.759	0.763
ibn_x1_0_msmt17_combineall_256x128_amsgrad_ep150_stp60_lr0	0.329	0.273	0.281	0.387	0.728	0.403
ms_d_c	0.389	0.457	0.407	0.531	0.685	0.650
ms_d_m	0.435	0.489	0.436	0.541	0.649	0.607
ms_m_c	0.426	0.407	0.492	0.641	0.746	0.726
x0_25_duke_256x128_amsgrad_ep180_stp80_lr0	0.370	0.500	0.430	0.535	0.755	0.693
x0_25_imagenet	0.517	0.514	0.634	0.611	0.766	0.749
x0_25_market_256x128_amsgrad_ep180_stp80_lr0	0.385	0.533	0.405	0.695	0.835	0.866
x0_25_msmt17_256x128_amsgrad_ep180_stp80_lr0	0.352	0.380	0.332	0.536	0.728	0.563
x0_25_msmt17_combineall_256x128_amsgrad_ep150_stp60_lr0	0.338	0.329	0.348	0.453	0.683	0.615
x0_5_duke_256x128_amsgrad_ep180_stp80_lr0	0.314	0.431	0.445	0.637	0.776	0.744
x0_5_imagenet	0.572	0.567	0.562	0.585	0.712	0.643
x0_5_market_256x128_amsgrad_ep180_stp80_lr0	0.302	0.442	0.412	0.741	0.885	0.869
x0_5_msmt17_256x128_amsgrad_ep180_stp80_lr0	0.405	0.402	0.388	0.621	0.711	0.663
x0_5_msmt17_combineall_256x128_amsgrad_ep150_stp60_lr0	0.276	0.355	0.265	0.565	0.639	0.478
x0_75_duke_256x128_amsgrad_ep150_stp60_lr0	0.341	0.517	0.453	0.644	0.764	0.701
x0_75_imagenet	0.577	0.524	0.604	0.688	0.756	0.778
x0_75_market_256x128_amsgrad_ep150_stp60_lr0	0.351	0.369	0.430	0.752	0.843	0.895
x0_75_msmt17_256x128_amsgrad_ep150_stp60_lr0	0.427	0.429	0.393	0.673	0.667	0.671
x0_75_msmt17_combineall_256x128_amsgrad_ep150_stp60_lr0	0.320	0.294	0.312	0.423	0.692	0.492
x1_0_duke_256x128_amsgrad_ep150_stp60_lr0	0.444	0.533	0.433	0.605	0.716	0.606
x1_0_imagenet	0.589	0.554	0.553	0.520	0.693	0.644
x1_0_market_256x128_amsgrad_ep150_stp60_lr0	0.349	0.514	0.506	0.746	0.882	0.801
x1_0_msmt17_256x128_amsgrad_ep150_stp60_lr0	0.438	0.438	0.447	0.526	0.655	0.638
x1_0_msmt17_combineall_256x128_amsgrad_ep150_stp60_lr0	0.341	0.285	0.265	0.476	0.686	0.504
4. resnet50
fc512_msmt_xent	0.821	0.808	0.779	0.835	0.859	0.890
msmt17_combineall_256x128_amsgrad_ep150_stp60_lr0	0.418	0.373	0.330	0.593	0.810	0.752