Collaborating Authors

Photorealistic Image Synthesis for Object Instance Detection Artificial Intelligence

We present an approach to synthesize highly photorealistic images of 3D object models, which we use to train a convolutional neural network for detecting the objects in real images. The proposed approach has three key ingredients: (1) 3D object models are rendered in 3D models of complete scenes with realistic materials and lighting, (2) plausible geometric configuration of objects and cameras in a scene is generated using physics simulations, and (3) high photorealism of the synthesized images achieved by physically based rendering. When trained on images synthesized by the proposed approach, the Faster R-CNN object detector achieves a 24% absolute improvement of mAP@.75IoU on Rutgers APC and 11% on LineMod-Occluded datasets, compared to a baseline where the training images are synthesized by rendering object models on top of random photographs. This work is a step towards being able to effectively train object detectors without capturing or annotating any real images. A dataset of 600K synthetic images with ground truth annotations for various computer vision tasks will be released on the project website:

Artificial Intelligence Can Unblur Pixelated Images


A team of researchers from the University of Texas at Austin and Cornell Tech have trained a software program to uncloak digitally-blurred or distorted images using deep learning, essentially teaching a computer to interpret a set of example data. Using this process, the team's software identified encrypted photographs with a 71% success rate. For context, the human success rate was 0.2%. The team purposely used an open source deep learning library to train their software, and their research exposes more weaknesses in the concept of online privacy. "The techniques we're using in this paper are very standard in image recognition, which is a disturbing thought," said Cornell Tech's Vitaly Shmatikov, pointing out that theirs was an "off-the-shelf, poor man's approach" to encrypted image recognition, and that a person or entity with bad intentions could do a lot of damage with this technology.

Kaggle Image Competitions! How to Deal with Large Datasets


When I have to deal with Huge image datasets, this is what I do. Working with image datasets in Kaggle competitions can be quite problematic, your computer could just freeze and don't care about you anymore. To stop this things from happening, I'm going to be sharing with you here the 5 Major Steps to work with Image datasets.

Satellite Images Can Harm the Poorest Citizens

The Atlantic - Technology

Mapping a city's buildings might seem like a simple task, one that could be easily automated by training a computer to read satellite photos. Because buildings are physically obvious facts out in the open that do not move around, they can be recorded by the satellites circling our planet. Computers can then "read" these satellite photographs, which are pixelated images like everyday photographs except that they carry more information about the light waves being reflected from various surfaces. That information can help determine the kind of building material and even plant species that appears in an image. Other patterns match up with predictable objects, like the straight lines of roads or the bends of rivers.



If you would like to apply a pre-trained model to a collection of input photos (without image pairs), please use --dataset_mode single and --model test options. For example, landscape painting - landscape photographs works much better than portrait painting - landscape photographs. For example, these might be pairs {label map, photo} or {bw image, color image}. A and B should each have their own subfolders train, val, test, etc.