Learning depth information from image is a crucial topic in computer vision. It is also a problem under the general topic Geometry learning This project target to explore and build machine learning model that can output depth or relative depth from input image either in a supervised or unsupervised manner.
Depth map prediction networks:
| Paper | Description |
|---|---|
| Learning Depth from Single Images with Deep Neural Network Embedding Focal Length | Fully supervised method considering varying focal length |
| Sparse-to-Dense: Depth Prediction from Sparse Depth Samples and a Single Image | Depth prediction with sparse depth samples augmentation |
| Depth Map Prediction from a Single Image using a Multi-Scale Deep Network | Coarse network + fine network (prior work for SOA on NYUv2) |
| Predicting Depth, Surface Normals and Semantic Labels with a Common Multi-Scale Convolutional Architecture | SOA model on NYUv2 in 2016 |
| Deeper Depth Prediction with Fully Convolutional Residual Networks | SOA model FRCN on NYUv2 using ResNet in 2016 |
| Deep Ordinal Regression Network for Monocular Depth Estimation | SOA model DORN on NYUv2 using ResNet in 2018 & 1st prize in Robust Vision Challange 2018 |
Global vs Local:
| Paper | Description |
|---|---|
| Non-local Neural Networks | Layer structure designed for spatial/time interactions or correlations |
| Large Kernel Matters -- Improve Semantic Segmentation by Global Convolutional Network | Proposed Global Convolutional Network for contradictory between classification and localization |
| Rethinking Atrous Convolution for Semantic Image Segmentation | Dilated convolution for dense prediction problem |
BerHu loss:
| Paper | Description |
|---|---|
| A unified approach to model selection and sparse recovery using regularized least squares | A smooth homotopy function between |
Image Models (texture):
| Paper | Description |
|---|---|
| Learning FRAME Models Using CNN Filters | Markov random field model for texture |
| A Theory of Generative ConvNet | Theory and Intuitions of Generative CNN |
| Generative Modeling of Convolutional Neural Networks | Generative modeling CNN |
| Filters, Random Fields and Maximum Entropy (FRAME): Towards a Unified Theory for Texture Modeling | FRAME model for texture |
| Learning Depth from Single Monocular Images Using Deep Convolutional Neural Fields | CRF Model with CNN for Depth Y given Image X |
| Deep Convolutional Neural Fields for Depth Estimation from a Single Image | Basically same work on CRF model |
| Multi-Scale Continuous CRFs as Sequential Deep Networks for Monocular Depth Estimation | CRF defined acorss multi-scale feature with attention gates inference by mean field approx |
| Structured Attention Guided Convolutional Neural Fields for Monocular Depth Estimation | Structured attention model jointly learned with CRF. Update: mean field approx |
Learning Energy based model (EBM):
| Paper | Description |
|---|---|
| On Learning Non-Convergent Short-Run MCMC Toward Energy-Based Model | Learning EBM using MCMC approach for approximating gradient |
| Energy-based Generative Adversarial Network | Energy based GAN, generator as a transformation sampler and discriminator as a energy function evaluator |
| Cooperative Learning of Energy-Based Model and Latent Variable Model via MCMC Teaching | Jointly learning an EBM with a latent variable model |
| Divergence Triangle for Joint Training of Generator Model, Energy-based Model, and Inference Model | Model 3 different joint distribution to avoid MCMC sampling |
| A Kernelized Stein Discrepancy for Goodness-of-fit Tests and Model Evaluation | Kernelized Stein Discrepancy (KSD) as a computable measure of discrepancy between a sample of an unnormalized distribution |
| Exponential Family Estimation via Adversarial Dynamics Embedding | "We consider the primal-dual view of the MLE for the kinectics augmented model, which naturally introduces an adversarial dual sampler." |
Implicit Learning Density with Score Matching (SM):
| Paper | Description |
|---|---|
| Generative Modeling by Estimating Gradients of the Data Distribution | Learning data generative score function using Score Matching and generate samples by Langevin Dynamics |
| A Connection Between Score Matching and Denoising Autoencoders | Denoising Score Matching (DSM) objective which can avoid caculate Hessian diagonal elements in Score Matching |
| Estimation of Non-Normalized Statistical Models by Score Matching | Score Matching(SM) |
| Information criteria for non-normalized models | Information criteria for noise contrastive estimation (NCE) and score matching. |
| Deep Energy Estimator Networks | Learning Energy function using Score Matching (SM) |
Satellite imagery data and biomass estimation:
| Paper | Description |
|---|---|
| Landsat-8: Science and product vision for terrestrial global change research | Landsat-8 Satellite imagery |
| An above-ground biomass map of African savannahs and woodlands at 25 m resolution derived from ALOS PALSAR | Estimation of above ground biomass over the whole Africa at a 25 m resolution. |
| Biomass estimation with high resolution satellite images: A case study of Quercus rotundifolia | Biomass estimation with high resolution satellite images: A case study of Quercus rotundifolia |
| Estimation and dynamics of above ground biomass with very high resolution satellite images in Pinus pinaster stands | Easily implemented in a GIS and a helpful tool in forest management and planning. |
| Landsat Imagery-Based Above Ground Biomass Estimation and Change Investigation Related to Human Activities | Landsat imagery and field data cooperated with a random forest regression approach were used to estimate spatiotemporal Above Ground Biomass (AGB) in Fuyang County, Zhejiang Province of East China. |
| Estimating Aboveground Biomass on Private Forest Using Sentinel-2 Imagery | Estimating Aboveground Biomass on Private Forest Using Sentinel-2 Imagery |
| Above-ground biomass estimation for Quercus rotundifolia using vegetation indices derived from high spatial resolution satellite images | The present study develops models to estimate and map above-ground biomass of Mediterranean Quercus rotundifolia stands using one QuickBird satellite image in pan-sharpened mode, with four multispectral bands (blue, green, red and near infrared) and a spatial resolution of 0.70 m. |