一种无监督人脸聚类方法(SOTA效果)

Learning to Cluster Faces by Infomap

Intorduction

采用了无监督方法infomap进行人脸聚类github，在公开数据集上MS-Celeb-1M、YouTube-Faces、DeepFashion获得较当前主流方法(如GCN人脸聚类等监督方法)同等或更优的效果.
通过faiss加速邻接边的构建，提高聚类速度，单批百万数据聚类仅需几分钟. 效果及性能测试详见下表.

Infomap Intorduction

About Infomap

About GCN Method

1. L-GCN: Linkage-based Face Clustering via Graph Convolution Network, CVPR 2019
2. GCN-D: Learning to Cluster Faces on an Affinity Graph, CVPR 2019 (Oral)
3. GCN-V+GCN-E: Learning to Cluster Faces via Confidence and Connectivity Estimation, CVPR 2020

Requirements

• Python >= 3.6
• sklearn
• infomap
• numpy
• faiss-gpu(or faiss-cpu)

Datasets

MS-Celeb-1M : part1_test (584K)、YouTube-Faces、DeepFashion
download

Run

python face-cluster-by-infomap

Results on part1_test (584K)

Method	Precision	Recall	F-score
Chinese Whispers (k=80, th=0.6, iters=20)	55.49	52.46	53.93
Approx Rank Order (k=80, th=0)	99.77	7.2	13.42
MiniBatchKmeans (ncluster=5000, bs=100)	45.48	80.98	58.25
KNN DBSCAN (k=80, th=0.7, eps=0.25, min=1)	95.25	52.79	67.93
FastHAC (dist=0.72, single)	92.07	57.28	70.63
DaskSpectral (ncluster=8573, affinity=‘rbf’)	78.75	66.59	72.16
CDP (single model, th=0.7)	80.19	70.47	75.02
L-GCN (k_at_hop=[200, 10], active_conn=10, step=0.6, maxsz=300)	74.38	83.51	78.68
GCN-D (2 prpsls)	95.41	67.77	79.25
GCN-D (5 prpsls)	94.62	72.59	82.15
GCN-D (8 prpsls)	94.23	79.69	86.35
GCN-D (20 prplss)	94.54	81.62	87.61
GCN-D + GCN-S (2 prpsls)	99.07	67.22	80.1
GCN-D + GCN-S (5 prpsls)	98.84	72.01	83.31
GCN-D + GCN-S (8 prpsls)	97.93	78.98	87.44
GCN-D + GCN-S (20 prpsls)	97.91	80.86	88.57
GCN-V	92.45	82.42	87.14
GCN-V + GCN-E	92.56	83.74	87.93
Infomap(ours)(k=50,min_sim=0.58)	95.50	92.51	93.98

[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-MrAsRZ9K-1606379388603)(./image/evaluate.png)]

Results on YouTube-Faces

Method	Pairwise F-score	BCubed F-score	NMI
Chinese Whispers (k=160, th=0.75, iters=20)	72.9	70.55	93.25
Approx Rank Order (k=200, th=0)	76.45	75.45	94.34
Kmeans (ncluster=1436)	67.86	75.77	93.99
KNN DBSCAN (k=160, th=0., eps=0.3, min=1)	91.35	89.34	97.52
FastHAC (dist=0.72, single)	93.07	87.98	97.19
GCN-D (4 prpsls)	94.44	91.33	97.97
Infomap(ours)(k=400,min_sim=0.56)	92.82	91.78	98.04

Results on DeepFashion

Method	Pairwise F-score	BCubed F-score	NMI
Chinese Whispers (k=5, th=0.7, iters=20)	31.22	53.25	89.8
Approx Rank Order (k=10, th=0)	25.04	52.77	88.71
Kmeans (ncluster=3991)	32.02	53.3	88.91
KNN DBSCAN (k=4, th=0., eps=0.1, min=2)	25.07	53.23	90.75
FastHAC (dist=0.4, single)	22.54	48.77	90.44
Meanshift (bandwidth=0.5)	31.61	56.73	89.29
Spectral (ncluster=3991, affinity=‘rbf’)	29.6	47.12	86.95
DaskSpectral (ncluster=3991, affinity=‘rbf’)	24.25	44.11	86.21
CDP (single model, k=2, th=0.5, maxsz=200)	28.28	57.83	90.93
L-GCN (k_at_hop=[5, 5], active_conn=5, step=0.5, maxsz=50)	30.7	60.13	90.67
GCN-D (2 prpsls)	29.14	59.09	89.48
GCN-D (8 prpsls)	32.52	57.52	89.54
GCN-D (20 prpsls)	33.25	56.83	89.36
GCN-V	33.59	59.41	90.88
GCN-V + GCN-E	38.47	60.06	90.5
Infomap(ours)(k=400,min_sim=0.88)	38.67	60.48	90.97

Time Consumes and Gpu Memory (k=50,min_sim=0.58)

Nodes	Edges	TimeCount	Gpu Memory
500000	16535263	160(s)	2745(MiB)
1000000	30206572	400(s)	3235(MiB)

Comments

• k值越大，TimeCount会变长，Gpu Memory会增加
• k值并非infomap的参数，仅为faiss构建knn构建时使用
• 数据量增加一倍，KNN构建时间为原来的4倍(实质为n*n的向量搜索)

References

• 最小熵原理（五）：“层层递进”之社区发现与聚类
• 人脸聚档主流方案