Re-Id Papers
1.Unsupervised Person Re-identification by Deep Learning Tracklet Association [ECCV 2018]
-
Keywords: Unsupervised, automatically generated person tracklet
-
Main idea: Tracklet Association Unsupervised Deep Learning (TAUDL): maximize cross-camera tracklet similarity and within-camera tracklet dissimilarity in an end-to-end deep learning framework.
2. Improving Person Re-identification by Attribute and Identity Learning [arVix 2017]
- Keywords: Attributes, multi-task
- Main idea: The author proposed a new attribute-person recognition (APR) network which combines the ID and attribute classification losses. 27 attributes for Market-1501 and 23 attributes for Duke are labeled.
3. Person Re-identification using CNN Features Learned from Combination of Attributes [ICPR 2016]
- Keywords: Fine-tuning, attributes
- Main idea: The author firstly conducts a fine-tuning (pre-trained AlexNet on ImageNet) on Pedestrian Attribute dataset (PETA), then applies metric learning (Cross View Quadratic Discriminant Analysis XQDA) on the target dataset.
4. Diversity Regularized Spatiotemporal Attention for Video-based Person-identification [CVPR 2018]
- Keywords: Video-based re-id, attention model
- Main idea: Spatial and temporal attentions are combined into a spatiotemporal attention models. Then a penalization term is used to regularize multiple redundant attentions.
5. Deep Association Learning for Unsupervised video Re-identification [BMVC 2018]
- Keywords: Unsupervised
- Main idea: Deep Association Learning (DAL): (1) local space-time consistency within each tracklet from the same camera view and (2) global cyclic ranking consistency between tracklet from disjoint camera views.
6. Unsupervised data association for Metric Learning in the context of Multi-shot Person Re-identification [AVSS 2016]
- Keywords: Metric learning
- Main idea: The main idea is to automatically label data with signatures represented by a set of multi-modal feature distributions (GMMs) for metric learning (Mahalanobis distance Learning) using KISSME algorithm.
7. Multi-shot Person Re-identification using Part Appearance Mixture [WACV 2017]
- Keywords: Metric learning
- Main idea: Like 6, firstly represent a person's appearance by a signature model, then compare distance between signatures.
8. Cross domain Residual Transfer Learning for Person Re-identification [WACV 2019]
- Keywords: Residual transfer learning
- Main idea: Turn fine tuning into 4 stages Residual Transfer Learning(RTL). From a pretrained VGG16, (a) Only train last layers like fine tuning, (b) add and train residual blocks, (c) train residual units and last layers ,(d) optional: retrain all layers.
| CUHK03 | Market-1501 | DukeMTMC | ||||
|---|---|---|---|---|---|---|
| R1 | mAP | R1 | mAP | R1 | mAP | |
| 1.TAUDL | 44.7 | 31.2 | 63.7 | 41.2 | 61.7 | 43.5 |
| 2.APR | 84.29 | 64.67 | 70.69 | 51.88 |
| PRID2011 | iLIDS-VID | MARS | |||||
|---|---|---|---|---|---|---|---|
| R1 | R5 | R1 | R5 | R1 | R5 | mAP | |
| 1.TAUDL | 49.4 | 78.7 | 26.7 | 51.3 | 43.8 | 59.9 | 29.1 |
| 4.SpaAtn+Q+TemAtn+Ind | 93.2 | 80.2 | 82.3 | ||||
| 5.DAL(ResNet50) | 85.3 | 97 | 56.9 | 80.6 | 46.8 | 63.9 | 21.4 |
| 6.MCM+KISSME | 64.3 | 86.1 | 40.3 | 69.9 | |||
| 7.PAM+LOMO+KISSME | 92.5 | 79.5 | |||||
| 8.RTL+XQDA | 92.1 | 78.7 | 67.9 |
| VIPeR | CUHK01 | PRID450S | GRID | |
|---|---|---|---|---|
| R1 | R1 | R1 | R1 | |
| 3.FT-CNN+XQDA | 42.5 | 46.8 | 58.2 | 25.2 |
| 3.FT-CNN+LOMO+XQDA | 52.1 | 62.3 | 71.5 | 29.1 |