level: CVPR author: Boqiang Xu (University of Chinese Academy of Sciences), Lingxiao He(AI Research of JD) date: 2020 keyword:

• Person re-identification

---

Paper: Black Re-ID

Summary

1. propose the study of the Black Re-ID problem and establish the first Black-reID dataset;
2. propose the head-shoulder adaptive attention network(HAA), which make use of the head-shoulder information to support person re-identification through the adaptive attention module, and can be integrated with the most current Re-ID framework and is end-to-end trainable;
3. both effective for Black Re-ID problem but also valid in similar clothing;

Research Objective

• Application Area:
• Purpose: retrieving the same person from overlapping cameras;

Proble Statement

previous work:

• most of pervious work extract features in terms of the attributes of clothing(color, texture), but it’s common for people to wear black clothes or be captured by surveillance systems in low light;
• Person Re-ID: treated as classification task, aiming at dividing person with same identity.
  • pose-based Re-ID, which uses an off-the-shelf pose estimator to extract the pose information for aligning body parts or generating person images; (network bigger and slower)
  • part-based Re-ID, which slices image or global feature into several horizontal grids, training individually and assembling for a discriminative person representation; (sensitive to the pose variations and occlusions)
  • leverage local information with attention maps, pay more attention to the regions of IOTs, and robust to the background clutter.
• Head-shoulder information: these features such as haircut, complexion or appearance offer abundant discriminative information;

Methods

• system overview:

【Adaptive Attention】 to determinate the global and local feature weights by distinguishing input types;

Evaluation

【Test One】 Accuracy comparison

【Test two】The Impact of Adaptive model

【Test three】Ablation Study of the GeM Pooling

【Test four】Performance comparison of global and head-shoulder features

Notes 去加强了解

• Code avaible: https://github.com/xbq1994/.

level: author: Jinlong Peng( Tencent), Fudan University, Nara Institute of Science and Technology date: 2020 keyword:

• MOT

Peng, J., Wang, C., Wan, F., Wu, Y., Wang, Y., Tai, Y., … & Fu, Y. (2020). Chained-Tracker: Chaining Paired Attentive Regression Results for End-to-End Joint Multiple-Object Detection and Tracking. arXiv preprint arXiv:2007.14557.

---

Paper: Chained-Tracker

Summary

1. propose an end-to-end online MOT model, to optimize object detection, feature extraction and data association simultaneously.
2. CTracker is the first method that converts the challenging data association problem to a pair-wise object detection problem;
3. design a joint attention module to highlight informative regions for box pair regression and the performance of CTracker is further improved;

Proble Statement

• existence of occlusions; object trajectory overlap; possibly challenging background
• propose a novel on-line model CTracker, which unifies object detection, feature extraction and data association into a single end-to-end model;

previous work:

• based on tracking-by-detection paradigm, contains three sequential subtasks: Object detection, feature extraction and data association; lead to local optima and more computation cost, discards the temporal relationships of adjacent frames
• Re-identification and attention: the former extracts more robust features for data association; the latter hep model focused; greatly increase the model complexity and computational cost
• **Detection-based MOT Methods: ** detection model and tracking model are completely independent, which is complex and time-consuming;
• Partially End-to-end MOT Methods:
• Attention-assistant MOT Methods:
  • Chu et. al[12] introduced a Spatial-Temporal Attention Mechanism to handle the tracking drift caused by the occlusion and interaction among targets.
  • [14] utilized an attention-based appearance model to solve the inter-object occlusion;

Methods

• system overview:

takes adjacent frame pairs as input to perform joint detection and tracking in a single regression model that simultaneously regress the paired bounding boxes for targets that appear in both of the two adjacent frames;

• end-to-end model using adjacent frame pair as input and generating the box pair representing the same target;
• convert the challenging corss-frame association problem into pair-wise object detection problem;

Notes 去加强了解

• code: github.com/pjl1995/CTracker