@alpizano
I have successfully extracted position data from the YoloV3 Network. You can see it in this
'cropped_short.mp4.txt' text file

The format of the data is as follows

| Point 1 x-value | Point 1 y-value | Point 2 x-value | Point 2 y-value | Class Index |
ConfidenceThreshold |
| -------------------- | -------------------- | -------------------- | -------------------- |----------------| --------------- |
|782 | 136 | 872 | 191 | 1 | 0.919052 |
|12 | 299 | 49 | 373 | 0 | 0.760167 |
|745 | 130 | 831 | 182 | 1 | 0.836964|
| 47 | 199 | 98 | 259 | 0 | 0.753106 |

Since Point 1 and Point 2 define a rectangle, fine the boxes center by finding the midpoint of the line defined by P1 and P2.

Then find the distance between this midpoint and the previously identified midpoint for the same class this will give you the relative velocity between frames of the video. Find the difference between these velocities to gain acceleration.

Find a way to define this into a file format then encode into polar coordinates. Or transform into polar then perform the above calculations. These values will then be fed into another network (i.e. predictive network) for training.

We have already initialized this project as a submodule, but it would be ideal to fork this project.

This is so we can remotely track 2 repositories in order to make custom changes and open source contributions.

We successfully trained the network for 1 clss, but we need to add support for more classes.
Classes needed:
1.Ball
2.Zero Fret

Optional
3. Wheel Center
4. Hands (to detect when ball has been released)

One of us should be able to use PyTorch to solve a classification problem. Code network using python and pytorch and push it into a separate folder in the learning exercise's folder in our repository

@mmassom96 and @alpizano

Please, both of you go over the video data collected from our specific roulette wheel.

1. Analyze the blur of the moving objects
2. Analyze time step (does it upon observation seem that time step is semi-consistent)
3. Analyze selected set of frames from each video, at your discretion.
4. Develop a document to report back to our meeting this Tuesday.

When building our network knowing what the ball looks like is useful, but we need to be able to do it with live video.

Efficiency (object detection which is as close to real-time as possible) is preferred. If there is too much lag or latency, it will cause issues later on.

We can use are prior object detection network as a building block.

Need to get images for multiple classes and use Amazon M Turk to annotate data for

1. Ball
2. 0 fret
3. Wheel center
4. Hand

Classes

Do the problem and push your proposed solution to a folder under your name in the
/examples/background_subtraction/

folder

It would be beneficial to have RAID 0 (possibly necessary) for training to avoid SSD damage.

@mmassom96

Please capture video data with the real sense camera from IEEE, so we can discuss it as an option at this Tuesday's meeting.

Then we can compare the data we got from the GoPro and the real sense to see which has better data.

Issue 281

Loss increases and model eventually locks up resources freezing.

Edit: Look at discussion in the link to ultralytics repo below.

When running detect.py, everything seems fine, but the video will not output to the specified file name. Need to look in yolov3/detect.py to see what is happening with the vid writer object.

Once resolved patch back to main repo. This will resolve this issue Issue 243 Ultralytics Yolo

@mmassom96
Clone this repo, and go into the yolov3 folder.

Check out the branch with git checkout br_develop

Follow the instructions here to learn how to train the network
Training Yolo

Your RTX2070 super has a lot of tensor cores which supports half-precision float models.
(16-bit floating-point precision vs 32-bit) which can speed up detection and training significantly.

Attempt to train network with cfg/yolov3-tiny.cfg or cfg/br_roulette_tiny3L.cfg with any data at half-point precision.

Evaluate performance and report back to the group your findings.

Do not hesitate to ask for help you need to.

@mmassom96

Compile a list of what you will need to construct the table, and select a height for the camera after we select between GoPro and RealSense.

@alpizano Have a skeleton frameworking of training classes and network classes.

Due on 10-08-2019 before meeting with Gray

Using PyTorch and python, begin coding a neural network which accepts and outputs these values.

Write a skeleton function to train the network.

Inputs

R-ball (Radius of ball from the center)
R-zero (Radius of zero from the center)
Theta-ball (Angular position of the ball in radians)
Theta-zero (Angular position of the zero)
W-ball (Radial speed of the ball in radians)
W-zero (Radial speed of wheel)
a - ball (angular acceleration of the ball)
a - zero (angular acceleration of the wheel)
T-time step between frames

Outputs

Theta final ball - final angular position of the ball
Theta final zero - final angular position of the zero

To implement object detection for smaller objects, we need to change hyperparameters and the configuratin of the network.

After doing some research and reaching out I have found these sources.

Low Feature Map Resolution

Talked to devlopers on darknet github and got response
Re: [pjreddie/darknet] How do I set the grid size when I use yolo v3? (#1042)

On changing grid size for more resolution.

This was the suggested
configuration file to enable a smaller grid size.

Transfer learning always wants to start out at yolov3-spp.pt, but this may not always be the optimal case.