Cleanup and no AWS code

[zubairahmed-ai]

Hi David

Just saw your code linked on Adrian's blog and I found the newest concurrent.futures.ThreadPoolExecutor that I didn't know existed after having tried Pool.map and Pool.starmap unsuccessfully so thanks for sharing your code

However going through it I find it hard to remove references to AWS and there are tons of commented code with textual comment which makes it difficult to adopt and run

If its not too much to ask could you please provide a copy of this version without the above with plan and simple object detection?

Thanks in advance

[00hello]

Hi Zubair. You're welcome w/re concurrent.futures.ThreadPoolExecutor.

I'll be releasing a new version of this "Node Lite" in the next few days with additional features but this "Node Lite" will always outsource its object detection to another machine. I'll be releasing a plain "Node" version in about a month and a half with all the same features except it'll do its object detection locally.

[zubairahmed-ai]

Thanks David looking forward to both

[zubairahmed-ai]

Hi Zubair. You're welcome w/re concurrent.futures.ThreadPoolExecutor.

I'll be releasing a new version of this "Node Lite" in the next few days with additional features but this "Node Lite" will always outsource its object detection to another machine. I'll be releasing a plain "Node" version in about a month and a half with all the same features except it'll do its object detection locally.

Hi David, any chance to getting the plain "Node" version?

[00hello]

Hi Zubair,

I hope to get the plain "Node" version up in the next few days. I had to rethink a lot of things over the holidays in order to get this to work the way I want it to.

I switched the system to a local database as I had planned to. However, I found that indexing global coordinates in a database efficiently is a problem scientists are still researching and it requires a special type of database structure usually referred to as a geospatial database. After some research I discovered S2 Geometry from Google. It's a lightweight and fast (Pokemon Go uses it) way of organizing geospatial data. And it's built for local key-value databases like the one we'll be using called LevelDB, which is also used by most cryptocurrencies and is very flexible as it allowed me to do something no other geospatial databases were able to give me. It let me add a timestamp to the locations of each object and store the data as arbitrary byte arrays, which will be very important later.

People were having trouble configuring the software --calibrating the node using the 3D GUI (the 3D map in the video on the website -> https://www.grassland.network/) of the world so the node knows what its position and orientation is in the real world so the objects it tracks are in their proper coordinates. So I had to fix some bugs in the GUI, make it easier to run and placed it inside the node so you don't need to download it separately.

The GUI can access the database too so you can use it to view the things your node has tracked. There's now a little analogue clock to the top right hand side of the interface with adjustable hour, minute and second hands; so as you rewind the clock you start to view things as they were in the past (I'll post a picture later).

Some of these features may seem superfluous but I found they were absolutely necessary to just do the basic things I wanted the software to do. And they also lay the foundation for what I expect to do with the software in the future.

[zubairahmed-ai]

Thanks
I'm looking for a version that can work in realtime with a video from a camera

[00hello]

Absolutely. That's what this is.

Its ability to work in real time depends largely on how powerful your GPU/CPU setup is, of course. For each object that passes in front of the camera you'll need at least one frame with the object in it to be sent to the object detector in order to identify it and find the precise bounding box and then the rest of the frames (either from before or after the detection) can be sent to a Mosse or Centroid tracker (both of which require much less computation) in order to track it. So in order to maintain real time speed depending on your GPU/CPU power, you can balance what percentage of your FPS you're sending to the object detector and what percentage you're sending to the tracker alone.

[zubairahmed-ai]

David

Thanks for explaining that, from Adrian's blog I noticed that he skips few frames then looks for new object to track, I'm looking forward to see your implementation of this

[00hello]

You're welcome.

You'll notice that this new version will use sockets for internal communication between the GUI and the node (Only the node can access its database; so the node will query its own database on the GUI's behalf) and that the object detection instance method also computes a single 256 bit hash (SHA-256) from specific hidden/internal activations (numpy arrays) of the neural network for each object detection. Which proves the model actually performed the computation. That's because the version that will come after this next one will use these socket API's to allow data sharing between nodes. So your node, my node and other nodes around the world can freely exchange data of the world their viewing with each other while knowing for sure that neither of us could have faked the data because of the hash.

This will really make the GUI map of the world come alive (especially after the model is upgraded from simple object detection to semantic segmentation ) because by exchanging information like this it will give all of us as data scientists and AI engineers data about the world that until now has only been available to large (and often predatory) data harvesting companies.

[zubairahmed-ai]

Hi David

I hope you're able to get the new version of your code out soon

The updates that you described seem great along with the improvements to the code that you planned to do

I just downloaded your code and cleaned up some commented unwanted code, hoping to integrate
your multiprocessing solution in my own code