How to get it working (22/09/2022)

[kliao1]

Context

The 2021 team didn't quite leave enough information on how to run this (oops).

TODO

• run the rosbags and demonstrate how to push the images through the pipeline
• show where the "pipeline" is modularised
• document the input and output of the pipeline

[kliao1]

We tried running depth_estimation out of the box after cloning it from the repo. There seems to be some problem with finding the xfeatures2d from opencv. I suspect opencv wasn't installed properly.

Instead, we found a Pre-2022 directory in the SSD in the Jetson. In this directory, we followed:
mur_stereo -> calibration -> learnopencv -> ObjectDetection-YOLO

Here, we used the following commands:

mkdir build && cd build
cmake ..
make
cd ..
./build/object_detection_yolo --image=<your-image-name>.jpg

While this enables us to test the training models, it is far from a solution since the 2022 goal is to integrate everything.

[kliao1]

To run both images, use the following commands:

./build/object_detection_yolo --left_image=hello.jpg --right_image=hello.jpg
./build/object_detection_yolo --left_vid=hello.mp4 --right_vid=hello.mp4

https://github.com/MURDriverless/depth_estimation/blob/fe37e3e994e1bb194074fd04b37a74f875d63ce9/object_detection_yolo.cpp#L772
From this section of the code outlines the epipolar geometry section that is used for computing the x,y, and z coordinates

https://github.com/MURDriverless/depth_estimation/blob/fe37e3e994e1bb194074fd04b37a74f875d63ce9/object_detection_yolo.cpp#L115
From this section outlines the camera calibration and setup

https://github.com/MURDriverless/depth_estimation/blob/fe37e3e994e1bb194074fd04b37a74f875d63ce9/object_detection_yolo.cpp#L176
From this section outlines the config options for how to run the stereo pipeline

https://github.com/MURDriverless/depth_estimation/blob/fe37e3e994e1bb194074fd04b37a74f875d63ce9/object_detection_yolo.cpp#L335
This net refers to darknet

[kliao1]

Here's some of my notes on running darknet
24.08_darknet_commands.odt

It should be useful if you want to retrain your YOLO model and test how it fares for different datasets. Not useful at all for integration but hey, one step at a time