ecosystem

pytorch >= 1.6

ignite

apex

ToDo

• 分割前景背景训练模型

• 复现fast-reid 各类trick

  • soft margin
  • non-local
  • gem pooling
  • circle loss
  • cosine lr scheduler 并绘图
  • 384×128

  

• 使用Cpython 提高eval速度

• 监督学习框架绘图

• 迁移学习框架绘图

• 增量学习框架绘图

• 搜索结果可视化

• 聚类效果绘图

• 特征绘图

• [ ]

parameter

direct

baseline:

• batch-size: 64
• instance: 4
• loss: softmax_triplet
• METRIC_LOSS_WEIGHT:1

market1501

train

backbone	id	tricks	mAP	rank1	mean
resnet50	01/02	baseline	85.8	94.2	90.10
resnet50	03	batch-size: 128	85.0	93.8	89.97
resnet50	04	loss: softmax_arcface_triplet	70.9	89.2	81.01
resnet50	05	METRIC_LOSS_WEIGHT: 2.0	86.5	94.1	90.30
resnet50	07	METRIC_LOSS_WEIGHT: 10.0	84.4	92.6	88.50
resnet50	08	METRIC_LOSS_WEIGHT: 5	85.6	93.9	89.70
resnet50	09	learning weight	86.7	93.9	90.36
se_resnet50	11	merge resnet50: 09	86.2	93.6	89.94
ibn_a_resnet50	12	merge resnet50: 09	88.3	95.0	91.68
ibn_b_resnet50	13	merge resnet50: 09	83.6	92.8	88.17
se_ibn_a_resnet50	14	merge resnet50: 09	87.9	94.5	91.21
se_ibn_b_resnet50	15	merge resnet50: 09	82.6	92.6	87.63
resnet50	16	merge resnet50: 05+09	86.5	93.9	90.18
resnet50	17	merge resnet50: 16 $loss = a_p + (a_p - a_n + mergin)_+$	86.2	94.3	90.26
resnet101	01	learning weight	87.7	94.6	91.15
se_resnet101	02	merge resnet101: 01	87.5	94.3	90.98
ibn_a_resnet101	03	merge resnet101: 01	88.6	95	91.77
ibn_b_resnet101	04	merge resnet101: 01	83.6	92.8	88.22
se_ibn_a_resnet101	05	merge resnet101: 01	88.6	94.8	91.70
se_ibn_b_resnet101	06	merge resnet101: 01	83.7	93.5	88.57
resnext101_32x8d	01	learning weight	88.6	95.0	91.81
se_resnext101_32x8d	02	merge resnext101_32x8d: 01	88.3	95.0	91.66
ibn_a_resnext101_32x8d	03	merge resnext101_32x8d: 01	89.0	95.3	92.16
ibn_b_resnext101_32x8d	04	merge resnext101_32x8d: 01	84.1	93.1	88.61
se_ibn_a_resnext101_32x8d	05	merge resnext101_32x8d: 01	88.6	95.3	91.98
se_ibn_b_resnext101_32x8d	06	merge resnext101_32x8d: 01	83.4	92.7	88.29
test

using rerank

backbone	id	mAP	rank1	mean
resnet50	02	94.2	95.0	94.6
ibn_a_resnet50	12	94.4	95.6	94.97
ibn_a_resnet101	03	94.5	95.8	95.17
ibn_a_resnext101_32x8d	03	94.5	95.7	95.11

DukeMTMC-reID

train

backbone	id	tricks	mAP	rank1	mean
resnet50	01		76.4	86.7	81.58
ibn_a_resnet50	02	learning weight	75.4	86.0	80.65
ibn_a_resnet50	03	learning weight	76.8	87.9	82.33

msmt17

train

backbone	id	tricks	mAP	rank1	mean
resnet50	01	learning weight	46.2	64.8	55.6
ibn_a_resnet50	02	learning weight	51.4	69.8	60.6

Join train

market1501--DukeMTMC--msmt17

baseline and target

backbone	id	tricks	market1501 map/rank-1	dukemtmc map/rank-1	msmt17 map/rank-1
resnet50	01	Epoch: 520	86.8/94.6	78.6/88.0	48.9/67.5
ibn_a_resnet50	02	learning weight Epoch: 540

Continual learning

market1501--DukeMTMC

Learning without Forgetting (LwF)

• LOSS_TYPE: softmax_triplet
• ID_LOSS_WEIGHT: 1.0
• BASE_LR: 0.000035

backbone	id	tricks	market map/forget	duck map
ibn_b_50	01	ce_dist:10	47.0/-36.3	66.3
ibn_b_50	02	merge ibn_b_resnet50: 01 ID_LOSS_WEIGHT: 0.2	46.7/-36.6	65.0
ibn_b_50	03	tr_dist	48.6/-34.7	66.6
ibn_b_50	04	ce_dist:10 tr_dist	48.6/-34.7	66.6
ibn_b_50	05	LOSS_TYPE: triplet 30 epoch	47.1/-36.2	57.2
ibn_a_50		tr_dist
ibn_a_50		ce_dist:10
ibn_a_50		ce_dist:10 tr_dist

Encoder Based Lifelong Learning (EBLL)

after GAP

first-step: training market with autoencoder

backbone	id	tricks	ae-loss	market map	market rank-1
ibn_a_50	18	CODE_SIZE: 1024 LAMBDA: 0.01	ae: 0.0010 ae_l1: 0.0194 ae_l2: 0.0014	88	95
ibn_a_50	19	CODE_SIZE: 512 LAMBDA: 0.01	ae: 0.0010 ae_l1: 0.0104 ae_l2: 0.0014	88.1	94.9
ibn_a_50	20	CODE_SIZE: 1024 LAMBDA: 0.1	ae: 0.0029 ae_l1: 0.6125 ae_l2: 0.1796	87.7	94.7
ibn_a_50	21	CODE_SIZE: 1024 LAMBDA: 0.05	ae: 0.0010 ae_l1: 0.0915 ae_l2: 0.0014	88	94.6

second-step: training duck, using ibn_a_50: 19 weight.

CONTINUATION.LOSS_TYPE = 'tr_dist'

backbone	id	tricks	market map/forget	duck map
ibn_a_50	06	ae_bce_loss	42	72.8
ibn_a_50	07	ae_l1_bce_loss	41.8	72.9
ibn_a_50	08	ae_l2_bce_loss	41.6	73.0
ibn_a_50	09	bce: ae_loss ae_l1 ae_l2	41.8	72.7
ibn_a_50	10	ae_mse_loss	41.9	72.9
ibn_a_50	11	ae_l1_mse_loss	42.4	72.7
ibn_a_50	12	ae_l2_mse_loss	42.9	72.8
ibn_a_50	13	mse: ae_loss ae_l1 ae_l2	43.2	72.6

before GAP

Acknowledgements

person-reid

Bag of Tricks and A Strong ReID Baseline

michuanhaohao/reid-strong-baseline

Unsupervised Deep Embedding for Clustering Analysis

arxiv DeepClustering

A Discriminative Feature Learning Approach for Deep Face Recognition

centerloss

unsupervised

Self-similarity Grouping: A Simple Unsupervised Cross Domain Adaptation Approach for Person Re-identification(SSG)

OasisYang/SSG visualizing-dbscan-clustering

continual learning

Learning without Forgetting

Encoder Based Lifelong Learning

Pytorch-implementation-of-Encoder-Based-Lifelong-learning