Initial commit

Author: danjia21

LICENSE

@@ -1,2 +1,134 @@
-# DR-SPAAM-Detector
-DR-SPAAM detector. Code to be released.
+This repository contains the implementation of *DR-SPAAM: A Spatial-Attention and Auto-regressive Model for Person Detection in 2D Range Data* ([arXiv](https://arxiv.org/abs/2004.14079)).
+
+# DR-SPAAM Detector
+DR-SPAAM is a deep learning based person detector that detects persons in 2D range sequences obtained from a laser scanner.
+
+![](imgs/dets.gif)
+
+Although DR-SPAAM is a detector, it can generate simple tracklets, based on its spatial similarity module.
+
+![](imgs/tracks.gif)
+
+To interface with many robotic applications, an example ROS node is also included.
+The ROS node, `dr_spaam_ros` subscribes to the laser scan (`sensor_msgs/LaserScan`)
+and publishes detections as `geometry_msgs/PoseArray`.
+To use a different message defination, simply modify `_detections_to_ros_msg` 
+method in `src/dr_spaam_ros/dr_spaam_ros.py`.
+
+![](imgs/dr_spaam_ros.gif)
+
+## Quick Start
+We provide our complete training and eveluation code.
+If you would like to re-run experiments and make changes to our code, you can start out with the following scripts.
+
+First clone and install the repository.
+```
+git clone https://github.com/VisualComputingInstitute/DR-SPAAM-Detector.git
+cd dr_spaam
+python setup.py install
+```
+
+Download and put the [DROW dataset](https://github.com/VisualComputingInstitute/DROW) under `dr_spaam/data`.
+
+Download the checkpoints from the release section and put them under `dr_spaam/ckpts`.
+
+To measure the network inference time, run:
+```
+python bin/demo.py --time 
+```
+
+To visualize detections on an example sequence, run:
+```
+python bin/demo.py --dets
+```
+
+To visualize detections with tracklets, run:
+```
+python bin/demo.py --tracks
+```
+
+To train your own network, run:
+```
+python bin/train.py --cfg cfgs/dr_spaam.yaml
+```
+
+To evaluat a checkpoint, run:
+```
+python bin/eval.py --cfg cfgs/dr_spaam.yaml --ckpt ckpts/dr_spaam_e40.pth
+```
+(add `--val` to evaluate on the validation set)
+
+Integrating DR-SPAAM into other python projects is easy. 
+Here's a minimum example.
+```python
+import numpy as np
+from dr_spaam.detector import Detector
+
+# Detector class wraps up preprocessing, inference, and postprocessing for 
+# DR-SPAAM. Set `original_drow=True` to use the original DROW model instead. 
+# Use `stride` to skip scan points if faster inference speed is needed.
+ckpt = 'path_to_checkpoint'
+detector = Detector(ckpt, original_drow=False, gpu=True, stride=1)
+
+# set angular grid (this is only required once)
+ang_inc = np.radians(0.5)  # angular increment of the scanner
+num_pts = 450  # number of points in a scan
+detector.set_laser_spec(ang_inc, num_pts)
+
+# inference
+while True:
+    scan = np.random.rand(num_pts)  # scan is a 1D numpy array with positive values
+    dets_xy, dets_cls, instance_mask = detector(scan)  # get detection
+
+    # confidence threshold
+    cls_thresh = 0.4
+    cls_mask = dets_cls > cls_thresh
+    dets_xy = dets_xy[cls_mask]
+    dets_cls = dets_cls[cls_mask]
+```
+
+## ROS node
+We provide an example ROS node `dr_spaam_ros`. 
+First install `dr_spaam` to your python environment.
+Then compile the ROS package 
+```
+catkin build dr_spaam_ros
+```
+
+Modify the topics and the path to the pre-trained checkpoint at 
+`dr_spaam_ros/config/` and launch the node using
+```
+roslaunch dr_spaam_ros dr_spaam_ros.launch
+```
+
+Use the following code to convert a sequence from a DROW dataset into a rosbag
+```
+python scripts/drow_data_converter.py --seq <PATH_TO_SEQUENCE> --output drow.bag
+```
+
+Use RViz to visualize the inference result.
+A simple RViz config is located at `dr_spaam_ros/example.rviz`.
+
+## Inference time
+|        | AP<sub>0.3</sub> | AP<sub>0.5</sub> | FPS (RTX 2080 laptop) | FPS (Jetson AGX) |
+|--------|------------------|------------------|-----------------------|------------------|
+|DROW    | 0.638 | 0.659 | 95.8 | 24.8 |
+|DR-SPAAM| 0.707 | 0.723 | 87.3 | 22.6 |
+
+Note: In the original paper, we used a voting scheme for postprocessing.
+In the implementation here, we have replaced the voting with a non-maximum suppression,
+where two detections that are less than 0.5 m apart are considered as duplicates 
+and the less confident one is suppressed.
+Thus there is a mismatch between the numbers here and those shown in the paper.
+
+## Citation
+If you use DR-SPAAM in your project, please cite:
+```BibTeX
+@article{Jia2020DRSPAAM,
+  title =   {{DR-SPAAM: A Spatial-Attention and Auto-regressive
+              Model for Person Detection in 2D Range Data}},
+  author =  {Dan Jia and Alexander Hermans and Bastian Leibe},
+  journal = {arXiv:2004.14079},
+  year =    {2020}
+}
+```

README.md

@@ -0,0 +1,26 @@
+*~
+*__pycache__*
+*_ext*
+*.bag
+*.csv
+*.cu.o
+*.DS_Store
+*.ipynb_checkpoints
+*.pkl
+*.png
+*.pth
+*.pyc
+*.tfevents*
+*.yml
+
+.idea/
+.vscode/
+*.egg-info/
+build/
+ckpt/
+ckpts/
+data/
+dist/
+output/
+results/
+result_*/