4 ms·
I suspect this is a very difficult problem however just because of how varied individual plants can be... telling plants apart can be hard even for trained huma
by p1esk 8y ago
I suspect this is a very difficult problem however just because of how varied individual plants can be... telling plants apart can be hard even for trained humans.
Actually, it's not. Despite all the hype surrounding deep learning, tasks like telling similar looking plants apart is precisely what's it's good at. In fact, it will probably be better than most humans at this task. Moreover, it's simple to implement. You take off the shelf model (e.g. ResNet-50) trained on ImageNet, and finetune it on a dataset of specific plants you want it to recognize (~1000 examples of each plant). Then you can run that model on an iPhone (or even on much cheaper hardware)
A much harder problem is to make the robot move through the terrain without getting stuck. But even there, looking at what Boston Dynamics is doing, it seems like we are not that far off.
- mrfusion 8y agoWhy does it help to be trained on imagenet first? Also are there any tutorials to teach you how to do what you’re describing?
- troquerre 8y agoIt’s called transfer learning. The base layers of a model trained on imagenet are general enough that they’re helpful for the specific task of processing plants. The benefit is imagenet has much more training data which is why you’d train on that first and then fine-tune the model with your specific data. Fast.ai is a great beginner friendly resource.
- kaveh_h 8y agoRobots with legs are probably not an option, they are both costly in production and energy usage. Something more akin to https://farm.bot https://farm.bot but on a larger scale would probably work better. Maybe a hybrid solution like a CNC bot built on top a frame with wheels that could simply drive around a field and work on different quadrants would be a way forward.
- wlesieutre 8y agoYou could put it on a radial track built like center pivot irrigation systems. Maybe even on the same construction if you already have that.
- samatman 8y agoThis is a billion dollar business, if it can be made to ride on existing systems. Imagine being able to plant, weed, and harvest strawberries from a center pivot. Huge.
- wlesieutre 8y agoI'm betting you'd still need to install a new track designed for the purpose, but you might be able to clamp it on to the existing framework. Irrigation stuff is below and would be tough to move around, maybe the robots dangle off the side?
- mymythisisthis 8y agoSee my comment above, the original steam trackers were too heavy to put on a field, so they used a system of wires to pull machinery around. Two trackers drive in tandem down the sides of a field, with a wire between them. The machinery suspended from the wire.
- Iv 8y agoI am currently making a suspended robot to work on that exact problem. I hope to achieve what Farm Bot does but on a much higher scale. I live in Japan where most of the fields are ricefields. Wheels or legs are not really an option under these extremely muddy conditions.
- tomcam 8y agoWould love to hear more about this
- 8y ago
- aaron695 8y ago>> telling plants apart can be hard even for trained humans. > Actually, it's not. Just telling apart a small percent of plants in tight ecosystems is incredibly complex and takes experienced human experts for that tight ecosystem and I've never seen it done using hardware. All "apps" are bullshit vaperware, including ones which include human assistance. You can't even get, what's plant is that in my garden apps that actually work. Are you talking about a monoculture where you might have just 4 types of plants to differentiate between? This I might get. Or telling if it's not plant X.... also possible maybe.
- rdlecler1 8y agoIt’s a hard problem I know BASF has invested a lot of money working on this very problem.