@@ -11,12 +11,19 @@ English | [简体中文](README_cn.md)
...
@@ -11,12 +11,19 @@ English | [简体中文](README_cn.md)
-[Citations](#Citations)
-[Citations](#Citations)
## Introduction
## Introduction
PaddleDetection implements three multi-object tracking methods.
The current mainstream multi-objective tracking (MOT) algorithm is mainly composed of two parts: detection and embedding. Detection aims to detect the potential targets in each frame of the video. Embedding assigns and updates the detected target to the corresponding track (named ReID task). According to the different implementation of these two parts, it can be divided into **SDE** series and **JDE** series algorithm.
-[DeepSORT](https://arxiv.org/abs/1812.00442)(Deep Cosine Metric Learning SORT) extends the original [SORT](https://arxiv.org/abs/1703.07402)(Simple Online and Realtime Tracking) algorithm to integrate appearance information based on a deep appearance descriptor. It adds a CNN model to extract features in image of human part bounded by a detector. Here we use `JDE` as detection model to generate boxes, and select `PCBPyramid` as the ReID model. We also support loading the boxes from saved detection result files.
-[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding) is a fast and high-performance multiple-object tracker that learns the object detection task and appearance embedding task simutaneously in a shared neural network.
-**SDE** (Separate Detection and Embedding) is a kind of algorithm which completely separates Detection and Embedding. The most representative is **DeepSORT** algorithm. This design can make the system fit any kind of detectors without difference, and can be improved for each part separately. However, due to the series process, the speed is slow. Time-consuming is a great challenge in the construction of real-time MOT system.
-[FairMOT](https://arxiv.org/abs/2004.01888) focuses on accomplishing the detection and re-identification in a single network to improve the inference speed, presents a simple baseline which consists of two homogeneous branches to predict pixel-wise objectness scores and re-ID features. The achieved fairness between the two tasks allows FairMOT to obtain high levels of detection and tracking accuracy.
-**JDE** (Joint Detection and Embedding) is to learn detection and embedding simultaneously in a shared neural network, and set the loss function with a multi task learning approach. The representative algorithms are **JDE** and **FairMOT**. This design can achieve high-precision real-time MOT performance.
Paddledetection implements three MOT algorithms of these two series.
-[DeepSORT](https://arxiv.org/abs/1812.00442)(Deep Cosine Metric Learning SORT) extends the original [SORT](https://arxiv.org/abs/1703.07402)(Simple Online and Realtime Tracking) algorithm, it adds a CNN model to extract features in image of human part bounded by a detector. It integrates appearance information based on a deep appearance descriptor, and assigns and updates the detected targets to the existing corresponding trajectories like ReID task. The detection bboxes result required by DeepSORT can be generated by any detection model, and then the saved detection result file can be loaded for tracking. Here we select the `PCB + Pyramid ResNet101` model provided by [PaddleClas](https://github.com/PaddlePaddle/PaddleClas) as the ReID model.
-[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding) learns the object detection task and appearance embedding task simutaneously in a shared neural network. And the detection results and the corresponding embeddings are also outputed at the same time. JDE original paper is based on an Anchor Base detector YOLOv3 , adding a new ReID branch to learn embeddings. The training process is constructed as a multi-task learning problem, taking into account both accuracy and speed.
-[FairMOT](https://arxiv.org/abs/2004.01888) is based on an Anchor Free detector Centernet, which overcomes the problem of anchor and feature misalignment in anchor based detection framework. The fusion of deep and shallow features enables the detection and ReID tasks to obtain the required features respectively. It also uses low dimensional ReID features. FairMOT is a simple baseline composed of two homogeneous branches propose to predict the pixel level target score and ReID features. It achieves the fairness between the two tasks and obtains a higher level of real-time MOT performance.
DeepSORT does not need to train, only used for evaluation. Before DeepSORT evaluation, you should get detection results by a detection model first, here we use JDE, and then prepare them like this:
DeepSORT does not need to train on MOT dataset, only used for evaluation. Before DeepSORT evaluation, you should get detection results by a detection model first, and then prepare them like this:
```
```
det_results_dir
det_results_dir
|——————MOT16-02.txt
|——————MOT16-02.txt
...
@@ -67,18 +69,44 @@ det_results_dir
...
@@ -67,18 +69,44 @@ det_results_dir
|——————MOT16-11.txt
|——————MOT16-11.txt
|——————MOT16-13.txt
|——————MOT16-13.txt
```
```
For MOT16 dataset, you can download the det_results_dir.zip of one detection model provided by PaddleDetection：
JDE used 8 GPUs for training and mini-batch size as 4 on each GPU, and trained for 30 epoches.
### FairMOT Results on MOT-16 Training Set
### FairMOT Results on MOT-16 Training Set
...
@@ -98,10 +126,11 @@ Each txt is the detection result of all the pictures extracted from each video,
...
@@ -98,10 +126,11 @@ Each txt is the detection result of all the pictures extracted from each video,
**Notes:**
**Notes:**
FairMOT used 8 GPUs for training and mini-batch size as 6 on each GPU, and trained for 30 epoches.
FairMOT used 8 GPUs for training and mini-batch size as 6 on each GPU, and trained for 30 epoches.
## Dataset Preparation
## Dataset Preparation
### MOT Dataset
### MOT Dataset
PaddleDetection use the same training data as [JDE](https://github.com/Zhongdao/Towards-Realtime-MOT) and [FairMOT](https://github.com/ifzhang/FairMOT). Please refer to [PrepareMOTDataSet](../../docs/tutorials/PrepareMOTDataSet.md) to download and prepare all the training data including **Caltech Pedestrian, CityPersons, CUHK-SYSU, PRW, ETHZ, MOT17 and MOT16**. **MOT15 and MOT20** can also be downloaded from the official webpage of MOT challenge. If you want to use these datasets, please **follow their licenses**.
PaddleDetection use the same training data as [JDE](https://github.com/Zhongdao/Towards-Realtime-MOT) and [FairMOT](https://github.com/ifzhang/FairMOT). Please refer to [PrepareMOTDataSet](../../docs/tutorials/PrepareMOTDataSet.md) to download and prepare all the training data including **Caltech Pedestrian, CityPersons, CUHK-SYSU, PRW, ETHZ, MOT17 and MOT16**. The former six are used as the mixed dataset for training, and MOT16 are used as the evaluation dataset. In addition, you can use **MOT15 and MOT20** for finetune. All pedestrians in these datasets have detection bbox labels and some have ID labels. If you want to use these datasets, please **follow their licenses**.
### Data Format
### Data Format
These several relevant datasets have the following structure:
These several relevant datasets have the following structure:
The default evaluation dataset is MOT-16 Train Set. If you want to change the evaluation dataset, please refer to the following code and modify `configs/datasets/mot.yml`：
```
EvalMOTDataset:
!MOTImageFolder
task: MOT17_train
dataset_dir: dataset/mot
data_root: MOT17/images/train
keep_ori_im: False # set True if save visualization images or video
-[DeepSORT](https://arxiv.org/abs/1812.00442)(Deep Cosine Metric Learning SORT) 扩展了原有的[SORT](https://arxiv.org/abs/1703.07402)(Simple Online and Realtime Tracking)算法，增加了一个CNN模型用于在检测器限定的人体部分图像中提取特征，在深度外观描述的基础上整合外观信息。
-[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding)是一个快速高性能多目标跟踪器，它是在共享神经网络中同时学习目标检测任务和外观嵌入任务的。
-SDE(Separate Detection and Embedding)这类算法完全分离Detection和Embedding两个环节，最具代表性的就是**DeepSORT**算法。这样的设计可以使系统无差别的适配各类检测器，可以针对两个部分分别调优，但由于流程上是串联的导致速度慢耗时较长，在构建实时MOT系统中面临较大挑战。
- JDE(Joint Detection and Embedding)这类算法完是在一个共享神经网络中同时学习Detection和Embedding，使用一个多任务学习的思路设置损失函数。代表性的算法有**JDE**和**FairMOT**。这样的设计兼顾精度和速度，可以实现高精度的实时多目标跟踪。
-[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding)是在一个单一的共享神经网络中同时学习目标检测任务和embedding任务，并同时输出检测结果和对应的外观embedding匹配的算法。JDE原论文是基于Anchor Base的YOLOv3检测器新增加一个ReID分支学习embedding，训练过程被构建为一个多任务联合学习问题，兼顾精度和速度。
[DeepSORT](https://arxiv.org/abs/1812.00442)(Deep Cosine Metric Learning SORT) extends the original [SORT](https://arxiv.org/abs/1703.07402)(Simple Online and Realtime Tracking) algorithm to integrate appearance information based on a deep appearance descriptor. It adds a CNN model to extract features in image of human part bounded by a detector. Here we use `JDE` as detection model to generate boxes, and select `PCBPyramid` as the ReID model. We also support loading the boxes from saved detection result files.
[DeepSORT](https://arxiv.org/abs/1812.00442)(Deep Cosine Metric Learning SORT) extends the original [SORT](https://arxiv.org/abs/1703.07402)(Simple Online and Realtime Tracking) algorithm, it adds a CNN model to extract features in image of human part bounded by a detector. It integrates appearance information based on a deep appearance descriptor, and assigns and updates the detected targets to the existing corresponding trajectories like ReID task. The detection bboxes result required by DeepSORT can be generated by any detection model, and then the saved detection result file can be loaded for tracking. Here we select the `PCB + Pyramid ResNet101` model provided by [PaddleClas](https://github.com/PaddlePaddle/PaddleClas) as the ReID model.
DeepSORT does not need to train on MOT dataset, only used for evaluation. Before DeepSORT evaluation, you should get detection results by a detection model first, here we use JDE, and then prepare them like this:
DeepSORT does not need to train on MOT dataset, only used for evaluation. Before DeepSORT evaluation, you should get detection results by a detection model first, and then prepare them like this:
```
```
det_results_dir
det_results_dir
|——————MOT16-02.txt
|——————MOT16-02.txt
...
@@ -31,9 +37,13 @@ det_results_dir
...
@@ -31,9 +37,13 @@ det_results_dir
|——————MOT16-11.txt
|——————MOT16-11.txt
|——————MOT16-13.txt
|——————MOT16-13.txt
```
```
For MOT16 dataset, you can download the det_results_dir.zip provided by PaddleDetection：
[FairMOT](https://arxiv.org/abs/2004.01888) focuses on accomplishing the detection and re-identification in a single network to improve the inference speed, presents a simple baseline which consists of two homogeneous branches to predict pixel-wise objectness scores and re-ID features. The achieved fairness between the two tasks allows FairMOT to obtain high levels of detection and tracking accuracy.
[FairMOT](https://arxiv.org/abs/2004.01888) is based on an Anchor Free detector Centernet, which overcomes the problem of anchor and feature misalignment in anchor based detection framework. The fusion of deep and shallow features enables the detection and ReID tasks to obtain the required features respectively. It also uses low dimensional ReID features. FairMOT is a simple baseline composed of two homogeneous branches propose to predict the pixel level target score and ReID features. It achieves the fairness between the two tasks and obtains a higher level of real-time MOT performance.
The default evaluation dataset is MOT-16 Train Set. If you want to change the evaluation dataset, please refer to the following code and modify `configs/datasets/mot.yml`：
```
EvalMOTDataset:
!MOTImageFolder
task: MOT17_train
dataset_dir: dataset/mot
data_root: MOT17/images/train
keep_ori_im: False # set True if save visualization images or video
```
### 3. Inference
Inference a vidoe on single GPU with following command:
```bash
# inference on video and save a video
CUDA_VISIBLE_DEVICES=0 python tools/infer_mot.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml -oweights=https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608.pdparams --video_file={your video name}.mp4 --save_videos
```
**Notes:**
Please make sure that [ffmpeg](https://ffmpeg.org/ffmpeg.html) is installed first, on Linux(Ubuntu) platform you can directly install it by the following command:`apt-get update && apt-get install -y ffmpeg`.
[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding) is a fast and high-performance multiple-object tracker that learns the object detection task and appearance embedding task simutaneously in a shared neural network.
-[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding) learns the object detection task and appearance embedding task simutaneously in a shared neural network. And the detection results and the corresponding embeddings are also outputed at the same time. JDE original paper is based on an Anchor Base detector YOLOv3 , adding a new ReID branch to learn embeddings. The training process is constructed as a multi-task learning problem, taking into account both accuracy and speed.
Inference a vidoe on single GPU with following command:
```bash
# inference on video and save a video
CUDA_VISIBLE_DEVICES=0 python tools/infer_mot.py -c configs/mot/jde/jde_darknet53_30e_1088x608.yml -oweights=https://paddledet.bj.bcebos.com/models/mot/jde_darknet53_30e_1088x608.pdparams --video_file={your video name}.mp4 --save_videos
```
**Notes:**
Please make sure that [ffmpeg](https://ffmpeg.org/ffmpeg.html) is installed first, on Linux(Ubuntu) platform you can directly install it by the following command:`apt-get update && apt-get install -y ffmpeg`.
[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding)是一个快速高性能多目标跟踪器，它是在共享神经网络中同时学习目标检测任务和外观嵌入任务的。
[JDE](https://arxiv.org/abs/1909.12605)(Joint Detection and Embedding)是在一个单一的共享神经网络中同时学习目标检测任务和embedding任务，并同时输出检测结果和对应的外观embedding匹配的算法。JDE原论文是基于Anchor Base的YOLOv3检测器新增加一个ReID分支学习embedding，训练过程被构建为一个多任务联合学习问题，兼顾精度和速度。
