UditSinghParihar

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch import optim


import numpy as np
import math

UditSinghParihar

1. Attention is all you need, paper review by Yannic Kilcher.
2. Attention and Transformer Networks, lecture by Pascal Poupart.
3. Thomas Kipf blog post is the first place to start for GCN.
4. A good video tutorial is from Microsoft Research.
5. Online book on deep learning with implementation d2l.ai.

UditSinghParihar

1. Create gparted live usb using Tuxboot.
   1. In Tuxboot GUI, select GParted as distribution to be installed.
2. Follow these gparted instructions to add unallocated space into exiting partions.
   1. Boot into gparted live usb. Might need to change the booting order.
   2. To change the partition size, left click on the side of partition and then start dragging.
3. To directly boot into firmware setup utility or bios setup from command line in current session:
   1. sudo systemctl reboot --firmware-setup
   2. Taken from Ubuntu 16.04 - how can I disable Secure Boot?
4. Alterante hack to manipulate the partitions size is by creating a bootable pendrive of ubuntu and change existing disk sizes from gparted by booting in the pendrive. This works because of the fact that you cant change d

UditSinghParihar

1. Cuda and cudnn:
   1. Instructions
   2. Cuda Toolkit Archive
2. cuDNN Archive

UditSinghParihar

1. Experiment performed on Hierarchical Pose Graph Optimization.
2. In our formulation topological nodes come at higher level of hierarchy tree followed by dense pose graph nodes.
3. We are currently experimenting that whether just optimizing the pose graph of higher topological nodes is sufficient to recover the trajectory close to the groundtruth.
4. The aim is to exploit topological information to get faster pose graph optimization time and get good enough map for robot navigation from one topology to another topology.
5. Here we are assuming that our frontend would be able to detect loop closures at the intersection of topological nodes.

UditSinghParihar

1. Blog:
   1. Blog on Geoemtric Deep learning.
2. Thesis:
   1. Tulsiani Thesis
3. Paper:
   1. Unsupervised Learning of Depth and Ego-Motion from Video
   2. Multi-view Consistency as Supervisory Signal for Learning Shape and Pose Prediction
   3. Taking a Deeper Look at the Inverse Compositional Algorithm
4. DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion

UditSinghParihar

1. Video of topological classification without pose optimization.
2. Video of topological classification and pose optimization using Manhattan and Loop closure constraints from ICP, based on Nodes proposed by MLP.
   1. Batch optimization is done after every 40 poses, total 2800 poses were there.
   2. Yellow dashed line denotes the constraints between two poses.
3. Simultaneously robot's POV video and live manhattan graph generation can also be shown.

UditSinghParihar

1. Following are the results of noisy trajectory optmization in g2o using Cauchy Cost Function vs DCS Algorithm.

UditSinghParihar

1. Following are the results of dataset1's second half optimization, using manhattan constraints.

UditSinghParihar

1. Rotation rectified by aligning first 50 poses of noisy trajectory vs ground truth trajectory.
2. Previous results of optmization on star shaped trajectory with rotation effect