I guess if you want to avoid the branch obstacle at the edge of screen from being detected as left/right obstacles, you could infact include a small part of the middle tree trunk, when labeling the right/left obstacles, so that model will pickup obstacles that are only attached to the main trunk. Nevertheless, Fantastic video indeed 💯. Loved the fact that you came back to fully correct the model. Deep learning is often frustrating when you don't get it right, but the magic when it actually works makes up for everything. Keep doing more of these❤️
Thanks a mil @Arjun, agreed, could've isolated the middle section of the image and only treated that as the active portion! It was killing me that performance was not where it should've been...had to come back and fix it!
Very cool, Nick! Here's to another tool for reinforcement learning for games, yay! 1. I feel like LabelImg should be easier to use, particularly with 'pixelart' stuff like timberman. Label a few and then LabelImg should itself make guesses for the remaining unlabeled images, where you'd only have to confirm or decline the proposal, most via keyboard presses. The longer you do it the more confident it would get. Some sort of "self-escalation" 2. Also: I can't shake the feeling that a neural net should have an easier time (compared to real life footage, for sure!) to get the detection and bounding boxes right. I wouldn't say 'most', but still a lot of the frames it sees have almost identical pixel arrangements for the classes it tries to detect. The bobbing of the timberman barely qualifies as data augmentation, right!? 3. For an example like timberman, I suppose it could help a lot to, I dunno, use opencv to remove the background in all the labeled examples?! The objects we are trying to find have black contours. I'm sure opencv as something onboard to isolate that. Without the background there should be less variation, making it easier for the neural net. There is this example of one neural net, that got good at detecting some animal (boy, I'm blanking on the actual kind). Turned out the training data tended to show said animal against a snowy background, and the net basically picked up on that more than on the animal itself.
I love your channel and thank you so much for this video because I have been working on this and finally found a video that explains everything so well. Keep up the good work
# if the 's' key is selected, we are going to "select" a bounding # box to track if key == ord("s"): # select the bounding box of the object we want to track (make # sure you press ENTER or SPACE after selecting the ROI) box = cv2.selectROI("Frame", frame, fromCenter=False,showCrosshair=True) # create a new object tracker for the bounding box and add it to our multi-object tracker tracker = cv2.legacy.TrackerCSRT_create() trackers.add(tracker, frame, box)
I don't get it why the views of this channel is low. I will share this to my friends because it is awesome. I learn a lot. Sir, kindly can u also make a video on Unsupervised learning(Deep Boltzmann Machine and AutoEncoders) and also make some projects with that in TF. I only have some books which are very tough to understand. Saw some videos but most of them did not show how can we use them. Or you can suggest some videos where I can find some projects.
This is super cool! As always, huge huge thanks, much appreciated! This is extremely useful in so so many usecases, and you explain it in a way that i find easly adaptible to pretty much every scenario you can imagine :) Perfect! I hope the "Breakdown Boar" is happy, whereever it lives :D Seriously though, I enjoyed that section :) I play a lot of a game called "Oxygen Not Included" at the moment, it's a survival-ish colony-building type of game, and at first I thought that this wouldn't be all too interesting for a game like that, because making AIs that can play games is mostly awesome for more competitive game like sports games, racing games, RL, SC, games like that. But then I though it would be kinda cool to build 2 robot hands that can use mouse & keyboard and train them to play games like ONI mentioned above (or literally any game, really). Doesn't even need to be super good at the game at first, I think that getting an actual physical robot to play games with actual Mouse & Keyboard would be super cool in itself :D Tried googling to see if someone has worked on a project like that before, but couldn't find much (and the google results are diluted with tons of unrelated topics because the javascript guys decided to call one of their KBM scripts "robot" ...). Since you asked about those challenge videos: I must admit I don't mind the challenge part all that much, becaue I feel like it just introduces a bit of stress/pressure due to the soft time limit and I prefer more calm/relaxed types of videos generally. I feel that too much "stress" makes it harder to learn, which is the ultimate goal. But I'm atistic, so that probably has influence on it aswell. It's not really a huge deal though ultimately and I wouldn't want to be the showstopper for those that do enjoy the challenge aspect, so I'm kinda neutral on that topic I'd say :) Would also definitely love to see this integrated into a larger project. Chaining together multiple pipelines ultimately isn't that hard, but I think it's still a lot of fun and there's always valuable info in your tutorials :) Especially when it comes to optimizing one pipeline with the following already in mind, if you know what i mean. Again, chaining them together is not difficult usually, but there are differences to working with them in isolation, different problems that can arise, different optimizations that are possible so that one model works better in conjunction with another, etc etc. Anyways, thanks again! Super awesome!
Oxygen not included looks amazing btw! Yeah, I think the challenge component is more to keep me honest otherwise I can spend weeks making a tutorial. There's so many half finished ones sitting on my desktop that if I had a little pressure they could be knocked off in no time. The desktop rl pipeline that I've got coming soon should be able to do what you described! I'm going to tidy it up this morning and look to get it out soon!
Thought about this some more in the last minute. I think for completely new topics, I'm still going to take it slow as I really want to set a good foundation for the channel e.g. super resolution I'm going to do a full blown breakdown.
@@NicholasRenotte It is a great game and also includes a lot of automation and that sort of stuff, which is probably one huge reason for why i personally really enjoy it :) And if that little bit of pressure is what keeps you motivated and running, then I'm totally happy with it :) I want you to enjoy this at least as much as I do, that's ultimately one of the most important things. As you say, it's also about finding the right mix, going slow where it's critical to be very very clear and in-depth and going faster when you need it or when you go over topics that for example already have been covered in your videos before.
hi sir this was a great video. can u make another on integrating this with the timberman game for actually making the python to play timberman based on the object detected from this script.
If you have not played around with opencv rtacker and selectROI i would take a look. You can get a lot of data by selecting a region and setting this region on the tracker.
@Nick Request you to put a tutorial on how to implement any object tracking (either deep sort or any other good frameworks) and integrate it with different object detection frameworks (YOLO v5 or TFOD ). Please 🥺 ( i requested this before as well ) 😞
This YOLOv5/PyTorch thing seems to have a shit-ton of memory leaks. Trying to train it is consuming all CUDA memory and system memory. I've had to reboot constantly.
Good afternoon! It seems to me that in chapter No. 6 you should send “cropped_region” instead of “corrected_colors” to the model, then the result will improve. Tell me if I'm wrong.😊
Hey Nicholas, I just watched your Multiperson Pose estimation which uses Movenet from Tensorflow. I have a request, is it possible for you to make a tutorial of Skeleton detection/ pose estimation using CenterNet. It would be a great use to me!! Or if you may provide a github for that. I am not able to find a proper working for the CenterNet Model. Thanks in advance.
My train.py step is not working as it is supposed to be. The error says the paging file size is too small. I've tried a lot of things to overcome this but nothing works. PLease help!!!!!!!!!!
Nice video but i wouldn't mind you not rushing it just to get to be 30 minutes, just accelerate or skips parts that repetitive such as labeling for instance :)
lol i mean i like your videos and all but this title is a huge bs, its not in 30 minutes, its not a bot and its not even working really still ig i learned something
