- 16
- 112 672
Fredrik Gustafsson
Приєднався 23 тра 2013
fregu856.com
Towards Accurate and Reliable Deep Regression Models | PhD Defense
www.fregu856.com
Thesis: www.fregu856.com/files/thesis.pdf
Slides: www.fregu856.com/files/defense_slides_handout.pdf
Abstract:
Regression is a fundamental machine learning task with many important applications within computer vision and other domains. In general, it entails predicting continuous targets from given inputs. Deep learning has become the dominant paradigm within machine learning in recent years, and a wide variety of different techniques have been employed to solve regression problems using deep models. There is however no broad consensus on how deep regression models should be constructed for best possible accuracy, or how the uncertainty in their predictions should be represented and estimated.
These open questions are studied in this thesis, aiming to help take steps towards an ultimate goal of developing deep regression models which are both accurate and reliable enough for real-world deployment within medical applications and other safety-critical domains.
The first main contribution of the thesis is the formulation and development of energy-based probabilistic regression. This is a general and conceptually simple regression framework with a clear probabilistic interpretation, using energy-based models to represent the true conditional target distribution. The framework is applied to a number of regression problems and demonstrates particularly strong performance for 2D bounding box regression, improving the state-of-the-art when applied to the task of visual tracking.
The second main contribution is a critical evaluation of various uncertainty estimation methods. A general introduction to the problem of estimating the predictive uncertainty of deep models is first provided, together with an extensive comparison of the two popular methods ensembling and MC-dropout. A number of regression uncertainty estimation methods are then further evaluated, specifically examining their reliability under real-world distribution shifts. This evaluation uncovers important limitations of current methods and serves as a challenge to the research community. It demonstrates that more work is required in order to develop truly reliable uncertainty estimation methods for regression.
Thesis: www.fregu856.com/files/thesis.pdf
Slides: www.fregu856.com/files/defense_slides_handout.pdf
Abstract:
Regression is a fundamental machine learning task with many important applications within computer vision and other domains. In general, it entails predicting continuous targets from given inputs. Deep learning has become the dominant paradigm within machine learning in recent years, and a wide variety of different techniques have been employed to solve regression problems using deep models. There is however no broad consensus on how deep regression models should be constructed for best possible accuracy, or how the uncertainty in their predictions should be represented and estimated.
These open questions are studied in this thesis, aiming to help take steps towards an ultimate goal of developing deep regression models which are both accurate and reliable enough for real-world deployment within medical applications and other safety-critical domains.
The first main contribution of the thesis is the formulation and development of energy-based probabilistic regression. This is a general and conceptually simple regression framework with a clear probabilistic interpretation, using energy-based models to represent the true conditional target distribution. The framework is applied to a number of regression problems and demonstrates particularly strong performance for 2D bounding box regression, improving the state-of-the-art when applied to the task of visual tracking.
The second main contribution is a critical evaluation of various uncertainty estimation methods. A general introduction to the problem of estimating the predictive uncertainty of deep models is first provided, together with an extensive comparison of the two popular methods ensembling and MC-dropout. A number of regression uncertainty estimation methods are then further evaluated, specifically examining their reliability under real-world distribution shifts. This evaluation uncovers important limitations of current methods and serves as a challenge to the research community. It demonstrates that more work is required in order to develop truly reliable uncertainty estimation methods for regression.
Переглядів: 193
Відео
Hur maskininlärning kan användas för att hjälpa läkare rädda liv
Переглядів 1,6 тис.2 роки тому
I denna video förklarar vi grundprincipen för hur maskininlärning fungerar samt hur detta kan användas för att hjälpa läkare upptäcka hjärtinfarkter. Förklaringen baseras helt på räta linjens ekvation och är särskilt anpassad för högstadieelever. Interaktiv version: educaora.com/@MachineLearningDoc Videon skapades som en del av doktorandkursen 'Using Maths and CS to do Social Good' på Uppsala u...
How Machine Learning Can Be Used to Help Doctors Save Lives
Переглядів 6642 роки тому
In this video, we explain the essence of how machine learning works and how it can be used to help doctors discover heart attacks from patient ECGs. The explanation is based on the concept of straight lines, and is specifically tailored to grade 7-9 students. Interactive version: educaora.com/@MachineLearningDoc Swedish version of the video: ua-cam.com/video/5ehdIBaElYA/v-deo.html The video was...
Accurate 3D Object Detection using Energy-Based Models | Qualitative Results
Переглядів 1 тис.3 роки тому
www.fregu856.com arXiv: arxiv.org/abs/2012.04634 Code: github.com/fregu856/ebms_3dod Project page: www.fregu856.com/publication/ebms_3dod/ Authors: Fredrik K. Gustafsson, Martin Danelljan, Thomas B. Schön Abstract: Accurate 3D object detection (3DOD) is crucial for safe navigation of complex environments by autonomous robots. Regressing accurate 3D bounding boxes in cluttered environments based...
How to Train Your Energy-Based Model for Regression | BMVC 2020 | 1.5 min Video
Переглядів 9743 роки тому
www.fregu856.com arXiv: arxiv.org/abs/2005.01698 Code: github.com/fregu856/ebms_regression Project page: www.fregu856.com/publication/ebms_regression/ Authors: Fredrik K. Gustafsson, Martin Danelljan, Radu Timofte, Thomas B. Schön Abstract: Energy-based models (EBMs) have become increasingly popular within computer vision in recent years. While they are commonly employed for generative image mo...
Energy-Based Models for Deep Probabilistic Regression | ECCV 2020 | 1 min Video
Переглядів 1,3 тис.4 роки тому
www.fregu856.com arXiv: arxiv.org/abs/1909.12297 Code: github.com/fregu856/ebms_regression Project page: www.fregu856.com/publication/dctd/ Authors: Fredrik K. Gustafsson, Martin Danelljan, Goutam Bhat, Thomas B. Schön Abstract: While deep learning-based classification is generally tackled using standardized approaches, a wide variety of techniques are employed for regression. In computer visio...
Evaluating Scalable Bayesian Deep Learning Methods for Robust Computer Vision | Qualitative Results
Переглядів 1,7 тис.5 років тому
Qualitative results for the paper: Evaluating Scalable Bayesian Deep Learning Methods for Robust Computer Vision, 2019. Fredrik K. Gustafsson, Martin Danelljan, Thomas B. Schön. arXiv: arxiv.org/abs/1906.01620 Code: github.com/fregu856/evaluating_bdl Project page: www.fregu856.com/publication/evaluating_bdl/ We propose a comprehensive evaluation framework for scalable epistemic uncertainty esti...
PyTorch implementation of DeepLabV3 | Semantic Segmentation for Autonomous Driving
Переглядів 16 тис.5 років тому
github.com/fregu856/deeplabv3 www.fregu856.com/ Segmentation is performed independently on each individual frame. 0:00 - 0:30: Cityscapes demo sequence 00. 0:30 - 1:25: Cityscapes demo sequence 01. 1:25 - 2:25: Cityscapes demo sequence 02. 2:25 - 4:24: Thn sequence (collected with a standard dash cam).
3D Object Detection for Autonomous Driving using Deep Learning (Master's Thesis Project)
Переглядів 17 тис.6 років тому
www.fregu856.com/ PyTorch implementation: github.com/fregu856/3DOD_thesis Frustum-PointNet trained on the KITTI dataset: 0:00. Image-Only model trained on the KITTI dataset: 1:58. Frustum-PointNet trained on the synthetic SYN dataset: 3:57. Extended Frustum-PointNet trained on the KITTI dataset: 5:55. Sequence order: KITTI test sequence 0011, 0002, 0007, 0001. The main video shows the estimated...
Semantic Segmentation for Autonomous Driving | TensorFlow Implementation of ENet
Переглядів 11 тис.6 років тому
www.fregu856.com/, github.com/fregu856/segmentation The results in the video can obviously be improved, but because of limited computing resources (personally funded Azure VM) I did not perform any further hyperparameter tuning. Segmentation is performed independently on each individual frame. 0:00 - 0:30: Cityscapes demo sequence 00 0:30 - 1:25: Cityscapes demo sequence 01 1:25 - 2:24: Citysca...
2D Object Detection for Autonomous Driving | TensorFlow Implementation of SqueezeDet
Переглядів 3,3 тис.6 років тому
www.fregu856.com/, github.com/fregu856/2D_detection The results in the video can obviously be improved, but because of limited computing resources (personally funded Azure VM) I did not perform any further hyperparameter tuning. Detection is performed independently on each individual frame. 0:00 - 0:41: KITTI training sequence 0001 0:41 - 1:27: KITTI testing sequence 0012 1:27 - 2:12: KITTI tes...
End-to-end obstacle avoidance system on an RC car
Переглядів 9536 років тому
www.fregu856.com/ I control the throttle, a neural network controls the steering. Work done as a part of a summer internship project.
BlueWind - 2D Adventure Game
Переглядів 2177 років тому
www.fregu856.com Final project in TDDC76: Programming and Datastructures.
Learning to Balance an Inverted Double Pendulum
Переглядів 10 тис.7 років тому
www.fregu856.com Final project in CS 229: Machine Learning.
Rasperino - Autonomous/Web Controlled Raspberry Pi & Arduino Robot
Переглядів 6 тис.7 років тому
www.fregu856.com Personal summer project. Code (~ 2500 lines) and info: github.com/dirac-hatt/Sommarprojekt16 Pictures from the build: goo.gl/photos/z78dbTpEZ5kwjqneA Main hardware: 1x Arduino Uno 1x Raspberry Pi 3 Model B 1x Raspberry Pi Camera Module v2 1x DFRobot 4WD Arduino Mobile Platform (www.robotshop.com/eu/en/dfrobot-4wd-arduino-mobile-platform.html) 2x L298 Dual H-Bridge DC Motor Cont...
Spider Pig - Autonomous Hexapod Robot (Bachelor Thesis Project, Linköping University)
Переглядів 41 тис.8 років тому
Spider Pig - Autonomous Hexapod Robot (Bachelor Thesis Project, Linköping University)
How did you manage to make it run the inference so fast? What GPU do you use?
this is boring asf
👍
really great
This is awesome 🔥🔥
This remind anyone else of the spider masterminds in DOOM..?
EI SENHORA HADES * GÁS * , os POBRES INIMIGOZINHOS das cajestudios Estão sendo bem tratados COM OS 220 WATS do TRON , AO CAMINHO ESTREITO DO ASFALTO E GARAGENS , QUANDO LANÇA OS FIOS ENGANTANDO NA CARNE IMUNDA DOS INIMIGOS DANDO MUITO CHOQUE ! hahahahahahahahahahahahahahahahahahahahahaha ! , resumindo em PARABOLA : * INIMIGO DO ROY , TEM QUE MORRER DEVAGAR * .
Looks good.
how much execution time will this model take on a single image if want to run on cpu
Am I the only one who expected to hear the "spider pig" theme sang by Homer Simpson?
Lidar technology it's beautiful ❤️
i love this, not afraid to have a utility shape, instead of just being sleek and nothing to it.
Hi sir ,I am currently working on this project could you please provide your email id ,it will be much help full,Thanks in advance.....
nice job!
Epic!! You should have way more views! Go and check out smzeus . c o m. It’s a very powerful tool, you can promote all of your social media profiles, UA-cam videos etc.
Did you just train it by recording yourself driving it and avoiding obstacles/driving straight?
you should learn to properly organize your work so that others can use it. Your GitHub repo is a mess and practically useless to anyone wanting to use your pre-trained models on their own set of images
Can it swing on a web?
does it have any drawbacks or limitations?
Wow So damn cool
Great job :) I checked your GitHub Repo, and I was wondering, which color correspond to which class. I think I got everything, except of the [81,0,81] (BGR) color (e.g. the "own car"). Is it possible, that this color corresponds to the remaining things? In cityscape, this would be black, right?
How does the robot understand that it has already stepped on the surface? mounted on the legs of the pressure sensors?
Nice Bro!!!
Open source code in website ? Im from indonesia, im build robo spider use arduino mega + shield mega and mpu 6050
Brilliant coding, well done.
This gives me a lot of motivations, Thanks! and great work!
Glad you liked it! :)
Well done!
Thanks! Glad you liked it. :)
Subscribed! Good job!
Thanks! Glad you liked it. :)
Good job ! My son and I just built a PhantomX MK3 and just got Phoenix code working today. really cool. We really want to learn how to make it autonomous like yours. Lots to learn lol. Thank you for the inspiration
Glad you found the video interesting, good luck! :)
The Bot remembers. XD
"thank's Fred...like your project
Brilliant robot, and would love to see this scaled up but with faster movement speeds! Is this a kit?
Hi Fredrik, Can we do this with Drone Image and LiDAR data? how to approach this problem. Can you share your report or code for this project? Thanks.
Hi Rishabh, I am working on the same topic. Can you share with me some good resources that you found?
Do you run the neural network on raspberry pi or on a server machine which communicates with it? Can the raspberry pi run the neural network fast enough?
kwea koi The network runs on a laptop, communicates over wifi. Didn't try running it on the RPi, but the network is very lightweight so it could probably work.
nice
really cool, fantasic, great work!
Thanks!
Hello Fredrik, Congratulations on your great work! I too am working on a similar problem. Would you please share your code and (or) your project report. Thanks, Sohom
I will link to the report as soon as it gets published, within a couple of weeks or so. Unfortunately, I might not be able to release the code. I'll update once I know for sure.
How to configure 32 channel servo controller v2.6 RTrobot + playstation 2 ?because i m confused grouping with the manual book. Can you lesson me step by step?
Great work. Did you train the network with cars and pedestrians at the same time?
Yes, I trained it to detect cars, pedestrians and cyclists, but it doesn't work nearly as well as for cars. I'd think the reason for this is that KITTI contains far more examples of cars than of pedestrians/cyclists.
have you tried to apply some class balancing techniques to solve the problem?
I am very interested your project. Can you share me source code?
Tin Huỳnh Đức I’m not sure where their code is, but here’s a paper describing the project: pdfs.semanticscholar.org/94ef/b6e4238c078ca0c443a1a538a52a54fa1e73.pdf
@@motherbear55 The paper is referencing to existing literature about model based control. So the assumption is, that the balancing controller is working with a model in the loop as well. But in the introduction it was promised, that a “simpler approach” over model based control was introduced. The name for this simpler approach is Deep Deterministic Policy Gradient, which is a model free learning algorithm.
how to program it?
I like the Homer Simpson reference.
Cool! Now, interface a spider or cockroach to drive that thing! instant insect exoskeleton!
IT MOVES!
3:40 Oh Sh#T! He escaped!! I hope mankind survives the wrath of Spider Pig
Awesome man! please share instruction about how do you make this hexapod robot,pls.tks
Jäklar vad grymt!!! Bra jobbat :D
Tack, ni får se till att göra något ännu coolare när det är er tur! :)
Hello Björn Runow, can you help me? I need your sketch, can you give me on your sketch? Rezaucheel@gmail.com
Great project man!