A brief conclusion on 'Occluded Person ReID'

This is a brief conclusion. I do this mainly for pushing myself to read more papers:)

About 'occluded person ReID'

This particular problem is first defined in the paper Occluded person re-identification (ICME 2018). It is quite simple by looking at the illustration as below.

Fig.1: images in the middle are occluded target persons; the goal is to retrieve a full-body person given a person with occlusions. red bdbox means the same one while green means different ones.

There are three challenges of occluded ReID mentioned in the paper.

1. Occlusions deteriorate global representations.
2. Matching corresponding parts need extra annotations.
3. The availability of datasets.

Here is the overview of the proposed approach.

Fig.2: Overview. An occlusion simulator and a multitask CNN.

• Occlusion Simulator (OS)

  
  
  

  Occlusion Simulator

The objective function:

OS objective

We need to make both and closer to . is the occluded image with the same label with .

• Multi-task Loss

Identification Loss:

Loss 1

OBC Loss:

Loss 2

Overall:

Overall Loss

In the experiments, the author mentioned four datasets.

1. Occluded-REID dataset: 2,000 images of 200 occluded persons, each of which has 5 full-body person images and 5 occluded person images.
   ( will be released)

2. P-DukeMTMC-reID and P-ETHZ datasets : will be released.

3. Partial-REID dataset: 900 images of 60 persons. Each identity has 5 full-body person images, 5 partial person images and 5 occluded person images.

The paper reports the classification, detection and ReID performances of the model on above four datasets.

I care more about this table. Haha...

image.png

The code is available at https://github.com/CaesarPan/Occluded-Person-ReID.

Some recent papers

1. Partial Person Re-identification with Alignment and Hallucination (ACCV2018)
2. SCPNet: Spatial-Channel Parallelism Network for Joint Holistic and Partial Person Re-identification (ACCV2018)
3. Partial person re-identification with two-stream network and reconstruction (Neurocomputing2019)
4. STNReID : Deep Convolutional Networks with Pairwise Spatial Transformer Networks for Partial Person Re-identification (arXiv 2019.3)
5. Deep Spatial Feature Reconstruction for Partial Person Re-identification:
   Alignment-free Approach (CVPR2018)
6. Perceive Where to Focus: Learning Visibility-aware Part-level Features
   for Partial Person Re-identification (CVPR2019)
7. Foreground-aware Pyramid Reconstruction for Alignment-free Occluded Person Re-identification (ICCV2019)
8. Pose-Guided Feature Alignment for Occluded Person Re-Identification (ICCV2019)

I will update this article in a few days.

Reference:

• Zhuo, Jiaxuan, et al. "Occluded person re-identification." 2018 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2018.
• He, Lingxiao, et al. "Deep spatial feature reconstruction for partial person re-identification: Alignment-free approach." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018.
• Sun, Yifan, et al. "Perceive Where to Focus: Learning Visibility-aware Part-level Features for Partial Person Re-identification." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.
• Fan, Xing, et al. "Scpnet: Spatial-channel parallelism network for joint holistic and partial person re-identification." Asian Conference on Computer Vision. Springer, Cham, 2018.
• Iodice, Sara, and Krystian Mikolajczyk. "Partial Person Re-identification with Alignment and Hallucination." Asian Conference on Computer Vision. Springer, Cham, 2018.
• Miao, Jiaxu, et al. "Pose-Guided Feature Alignment for Occluded Person Re-Identification." Proceedings of the IEEE International Conference on Computer Vision. 2019.
• Luo, Hao, et al. "STNReID: Deep Convolutional Networks with Pairwise Spatial Transformer Networks for Partial Person Re-identification." arXiv preprint arXiv:1903.07072 (2019).