Deep Learning Techniques for Monocular Depth Estimation

on YouTube ↗01h 30m

" summary

Covers three concrete systems with code and Colab notebooks: a transfer‑learning U‑Net using a DenseNet169 encoder with 42M parameters and a loss mixing pixel L1, gradient L1 and SSIM (trained up to 1M iterations) that sharpens edges; AdaBins: an EfficientNet‑B5 encoder + MiniViT adapter that predicts adaptive binning (per‑image bin widths), an AdaBins module of ~5.8M params and a 78M‑param model trained 25 epochs, batch 16; DPT vision‑transformer models (base/hybrid/large), with DPT hybrid 23% and large ~28% improvement over prior Mix datasets, plus pointers to DINOv2 pretraining.

AI-written summary · about

" chapters (12)

1. 0:00Monocular depth estimation intro
2. 1:26Today’s deep learning plan
3. 3:35Why monocular depth is hard
4. 9:18High-quality depth via transfer learning
5. 16:38Transfer learning demo in Colab
6. 21:42Adaptive bins for dynamic depth
7. 37:50Adaptive bins demo in Colab
8. 43:36Vision Transformers for dense predictions
9. 1:01:04DPT model demo in Colab
10. 1:08:36DINOv2 for vision tasks
11. 1:16:42Homework and resources
12. 1:29:07Q&A

" related talks

• 56mTutorial on Deep Learning IIRuslan Salakhutdinov · Foundations of Machine Learning Boot Camp 2017
• 37mWhat's new in Deep Learning?Kashif Rasul · PyData Berlin 2016
• 26mComputer Vision in the Era of Deep LearningGordan Kreković · Infobip Shift 2023
• 47mDeep Learning in Computer VisionKrzysztof Kudrynski, Błażej Kubiak · Voxxed Days CERN 2019
• 59mDeep Neural Networks for Video ApplicationsAlex Conway · PyConZA 2019
• 41mDeep Learning to Solve Challenging ProblemsJeff Dean · Google I/O 2019
• 01h 02mTutorial on Deep Learning IRuslan Salakhutdinov · Foundations of Machine Learning Boot Camp 2017
• 08mDeep Transfer Learning for Computer VisionDipanjan Sarkar, Sachin Dangayach · ML Conf EU 2020