tejus-gupta

State lattice-based planning

tl;dr

The kinodynamics constraints of the robot are encoded in the state lattice graph and any path in this graph is feasible. After constructing the graph, any graph search algorithm can be used for planning.

Algorithm

A robot's configuration space is usually discretized to reduce computational complexity of planning at the expense of completeness. However, it is difficult to search this space while satisfying the robot's differential constraints. State lattices are a special way of discretization of robot state space that ensures (by construction) that any path in the graph complies with the robot's constraints, thereby eliminating the need to consider them explicitly during planning.

tejus-gupta

Search and Heuristic

• A Note on Two Problems in Connexion with Graphs by Dijkstra
• A Formal Basis for the Heuristic Determination of Minimum Cost Paths by Hart
• On the complexity of admissible search algorithms by Martelli
• Heuristic Search Viewed as Path Finding in a Graph by Pohl
• R* Search by Likhachev abd Stentz
• Incremental A* by Koenig and Likhachev
• Lifelong planning A* by Koenig, Likhachev a

tejus-gupta

Semantic Segmentation

A Recurrent Neural Network-based Model for Semantic Segmentation
PyramidNet
PSPNet
Semantic Segmentation Models
Semantic Segmentation Models
FCN Large Kernel Matters -- Improve Semantic Segmentation by Global Convolutional Network
PointNet

tejus-gupta

Project Abstract

I proposed to

1. Add a package for image segmentation as part of JuliaImages with several algorithms -

• Thresholding - Otsu’s method and Adaptive thresholding
• K-means clustering
• Mean shift segmentation
• Watershed segmentation
• Felzenszwalb's efficient region merging algorithm
• Shi and Malik’s normalized graph-cut based segmentation

tejus-gupta

    if message_type == 'user_read':
        key,_ = data.split('|')

        highest_version = -1
        highest_version_value = ''

        #nodes = random.sample([(hash(key)+i)%N for i in range(R)], Q_r)
        #nodes = random.sample(get_next_live_inc(hash(key), R), Q_r)

        nodes = get_next_live_inc(hash(key), R)

tejus-gupta

<launch>
  <!-- 
  NOTE: You'll need to bring up something that publishes sensor data (see
  rosstage), something that publishes a map (see map_server), and something to
  visualize a costmap (see nav_view), to see things work.

  Also, on a real robot, you'd want to set the "use_sim_time" parameter to false, or just not set it.
  -->
  <param name="/use_sim_time" value="true"/>

tejus-gupta

to_remove = []

for class_idx in selected_classes:
	if np.sum(y_train[:, class_idx] < 0.5) < 5 or np.sum(y_train[:, class_idx] > 0.5) < 5:
		to_remove.append(class_idx)

for class_idx in to_remove:
	selected_classes.remove(class_idx)

print('Removed classes with too few examples')

tejus-gupta

import os
import sys
import yaml
import time
import shutil
import torch
import random
import argparse
import datetime
import numpy as np

tejus-gupta

import os
import sys
import yaml
import time
import shutil
import torch
import random
import argparse
import datetime
import numpy as np

tejus-gupta

mkdir train_set
cd train_set
wget https://gist.githubusercontent.com/tejus-gupta/3d4564e624cad79691706a5c1303f4c6/raw/3cafe4877f981e3f3c481727d0a90db519a4e95b/download.py
python download.py
unzip -qq masks.zip
unzip -qq train_data.zip
cd ..
git clone https://github.com/tejus-gupta/Segmentation
cd Segmentation
git checkout modelD