NVIDIA Research at CVPR 2019

NVIDIA Researchers will present 20 accepted papers and posters, eleven of them orals, at the annual Computer Vision and Pattern Recognition (CVPR) conference…

NVIDIA logo
NVIDIA
•
Nefi Alarcon
•20 min read•advanced•
--
•View Original
Computer VisionConvolutional Neural NetworksGenerative Adversarial NetworksNeural NetworksResNetUnsupervised Learning

Overview

The article discusses NVIDIA's contributions to the CVPR 2019 conference, highlighting 20 accepted papers and posters, including advancements in semantic image synthesis, video action detection, and person re-identification. It showcases innovative methodologies and frameworks that enhance computer vision tasks, emphasizing the importance of these research efforts in advancing the field.

What You'll Learn

1

How to implement spatially-adaptive normalization for image synthesis

2

Why spatio-temporal progressive learning improves video action detection

3

How to utilize generative models for person re-identification

4

When to apply video propagation techniques for semantic segmentation

5

How to leverage large-scale datasets for multi-camera vehicle tracking

Key Questions Answered

What is spatially-adaptive normalization and how does it improve image synthesis?
Spatially-adaptive normalization is a technique that modulates activations in normalization layers based on the input semantic layout, addressing the issue where traditional normalization layers wash away semantic information. This method enhances visual fidelity and alignment with input layouts, allowing for better user control over synthesized images.
How does the STEP framework enhance video action detection?
The STEP framework refines action proposals progressively from coarse to fine, allowing for better adherence to action movements over time. This approach results in superior detection performance, achieving a mean Average Precision (mAP) of 75.0% on the UCF101 dataset, demonstrating its effectiveness compared to traditional single-run detection methods.
What advancements does PlaneRCNN bring to 3D plane detection?
PlaneRCNN introduces a deep neural architecture that detects and reconstructs piecewise planar surfaces from a single RGB image. It employs a variant of Mask R-CNN and outperforms existing methods in plane detection, segmentation, and reconstruction metrics, significantly enhancing applications in Robotics and Augmented Reality.
What is the significance of the CityFlow dataset for vehicle tracking?
The CityFlow dataset is a city-scale benchmark featuring over 3 hours of synchronized HD videos from 40 cameras, making it the largest dataset for multi-target multi-camera vehicle tracking. It includes more than 200K annotated bounding boxes, facilitating advanced research and development in urban traffic optimization.

Key Statistics & Figures

mean Average Precision (mAP)
75.0%
Achieved on the UCF101 dataset using the STEP framework with 3 progressive steps.
mIoU on Cityscapes
83.5%
Achieved through the proposed video prediction-based methodology for semantic segmentation.
mIoU on KITTI semantic segmentation test set
72.8%
Surpassing the winning entry of the ROB challenge 2018.

Key Actionable Insights

1
Implementing spatially-adaptive normalization can significantly enhance the quality of generated images in computer vision tasks.
This technique allows for better preservation of semantic information during image synthesis, leading to more photorealistic outputs and improved user control over the synthesis process.
2
Utilizing the STEP framework for video action detection can lead to improved accuracy and efficiency in identifying actions in video data.
By refining action proposals progressively, this method can adapt to spatial displacements, making it more effective than traditional single-step detection methods.
3
Leveraging large-scale datasets like CityFlow is crucial for advancing multi-camera vehicle tracking technologies.
The extensive annotations and diverse scenarios provided by such datasets enable researchers to develop and benchmark more robust tracking algorithms.