Closed pkrush closed 6 years ago
After the Proof of Concept 2 milestone this is a logical next step: 1.) Test using the software on various Windows boxes and laptops.
I am going to need help with this one. I think a Vision Maker Meetup at Workshop88 is in order for this.
Coins have always been the first part of many. Sorting is the first task of many. I can see that this project will evolve quickly beyond coin sorting. Quicker than I thought it would.
I love working with small parts for at ton of reasons(another list for another day).
Other parts to work with:
I am digging through my screw bin for 10-24 screws right now. If trained could Caffe find 10-24 screws that are among 10-32 and 8-32 screws. Of course!
Detecting Arrowheads with drones & computer vision
It’s getting way easier to find stuff with computer software. For example you can train a program to understand the difference between a rock and an arrowhead in an image from a camera. I still think a human would be better unless camera was able to travel more ground, or “ground” was moving past the camera on a conveyor belt. Drones are really cool, but I have a feeling setting up lines for the camera to travel cable car style would be the way to go. Open ground and lots of light would be best.
Of course one could look for whatever they can see with their eyes: minerals, gems, fossils, and other relics. There would be an advantage in looking for tiny stuff. It’s easier to make the camera cover more ground and it could see small things much quicker than your eyes can.
What do you think? Do you see this happening? I do, it’s not half baked sci-fi. It wouldn’t be perfect, but it would speed up the hunt. I would love to build the equipment to do this. I am working with coins right now, but this tech (deep learning) is easy to adapt.
Out of curiosity what sorts of things would you look for? What area do you think people would be the most interested? Scanning for their lost keys?
Thanks!
Paul Krush GemHunt.com
Detect the same coin twice:
Yes! One could make Where’s George for coins.
I see this as a useful feature. People are going to look back at images from previous batches and really want a single coin. It would be a lot easier to do this then to look through a 5 gallon bucket it when into.
This would be cool for inventory systems as well. Virtual coin picking systems. Scan the junk bin and post it on a website. Then shipment time comes and the system can fine the same coin to ship out.
Before I tried deep learning I tried classical feature detectors such as SIFT. They could find the same coin, but had a lot of trouble finding the same design on a different coin. My bet is at a deep enough level of detail the nicks and wear on a coin is like a finger print positively ID coins twice. And what if you started finding worn coins with the same exact details? Counterfeits!
My current plans are to keep working on this coin sorting project till fall 2016. It’s based on the low end coin market. Finding & discovering new verities and errors. Low end grading. Low end auto rotate hobby inspection. Mass Coin Imaging. It’s all kind of related and the tools are the same. If the demand is there I want to try metal content detection with eddy currents, trained with XRF.
Pump live sorting video to Twitch. There are all sorts of image updates you could do. Live feed of the system. Live head shot. "CNN/ESPN" style feed of latest finds. OK, this is way out there, but would make for a great chat room.
It would be really nice if there was a way to automatically test camera setup performance. How the system performs as a whole(camera, lighting, focus,etc).
Make a many bin sorting conveyor on a 72” x 30” table
I don’t know how people would really want to do this, but it can be done for pretty cheap.
Out of curiosity I did a mock up where one solenoid knocks a coin into ether 2 boxes or chutes. The chutes could go to bags in the floor.
You could go nuts with this. One could use more levels and smaller boxes. For a demo I think clear boxes would be neat.
The boxes would be 4”x4”x8” boxes cut in half on a small angle. You can buy this for about 30 cents each, but they happen to be the same size as penny boxes.
One alternative to save space and make the conveyor use less wood the conveyor frame could sit on top of the boxes. This would make it hard to clean out the boxes, but they would be locked in place.
OK so that post above would be more interesting if each bin was a person. A high speed in person low value coin trading system or call it the junk bin swapping system See SortNSwap.com. Say you have 8 people around a 60" round table and conveyors shuttle coins to and from each person. The conveyors would be under glass so the table would be mostly usable. I though of this years ago. One computer can control it with displays mounted in the middle and above the table so people could still see each other. Each person could get a $1 keypad. A low cost system. Of course today it could just be smart phone based... For more fun this could incorpate gambling. Not just coins could be moved. Cards, poker chips, gem bags, anything small.
Label the facts of wear and the general grade? It would be nice to label coins down to yes/no questions. It would be nice to see how the labels and the grade correlate.
There is clear demand for a coin sorting and inspection tool, but it’s not just people interested in coins that would use this. Since everyone one understands what coins are, coins a great "small part" to demonstrate what a low cost system like this can do. It's also a great project to help people learn computer vision.
Beyond coins, the overall goal is a open low cost system / kit that can:
With such a system users would train the network for their own parts.