sagniklp

'''
This code was used to capture documents for the Doc3D dataset
For a document it captures the point cloud and the segmentation map
This code uses python (2.7) and pyrealsense2 (2.16) and may be incompatiable with newer versions of pyrealsense2
Related publication: DewarpNet: Single-Image Document Unwarping With Stacked 3D and 2D Regression Networks (ICCV 2019)
Contributed by: Ke Ma and Sagnik Das (Spring 2019)
Correspndence: sadas@cs.stonybrook.edu
Acknowledgemnts: This code uses snippets from pyrealsense2 tutorials
'''

sagniklp

sagniklp

Half Fuller Buttons

I really liked the buttons on this behance post, so I wanted to try and make some for my site. They are wicked cool.

https://www.behance.net/gallery/27689051/Paktor-App

A Pen by Sagnik Das on CodePen.

License.

sagniklp

# Helpful Tips: https://stackoverflow.com/questions/9567882/sobel-filter-kernel-of-large-size/
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import numpy as np
np.set_printoptions(threshold=np.nan)
from math import exp
import cv2
import matplotlib.pyplot as plt

sagniklp

Motivations:

• Predicting 3D shape of a deformable non-rigid surface from a single image.
• Original paradigm is SfT (Shape-from-Template) requires a reference template image of the surface for which the 3D geometry is known, and a set of 3D-to-2D point correspondences or a mapping between this template and the input image. This method struggles with low-textured surfaces.

Goals:

• This methods idea builds upon designing a method capable of predicting geometrically consistent 3D shape without relying on point correspondences.
• Robustness to lack of texture, large occlusion, corrupted object boundaries and varying lighting conditions.
• An end-to-end fully differentiable network.

Contributions:

sagniklp

import os
import numpy as np
import tqdm

inp_pc_path='./simple.xyz'  #input point cloud
out_pc_path='./simple_org.xyz'  #output point cloud

points=[]
org_points=[]
with open(inp_pc_path,'r') as f:

sagniklp

sagniklp

Motivations:

• Image-to-Image translation is a classical problem in image processing, graphics and vision.
• Requires manual effort to design "what to minimize" (loss), also being "careful what we wish for".
• Same architecture and objective of mapping two images with different training data without any manual mapping functions.

Being a traditionalist, what we could do?

• Take a CNN, find what loss works with your training data
• Maybe, just minimize the Euclidean distance of the ground-truth and predicted pixels. Result: Blurry Images

Goals: