Learning Notes - Research papers - Heatmaps

Paper #1 - Revisiting Perspective Information for Efficient Crowd Counting
Key Notes

• Perspective-aware convolutional neural network (PACNN)
• Estimate crowd counts via the detection of each individual pedestrian
• Crowd counting is casted as estimating a continuous density function

Detection-based methods

• Represent the crowd as a group of detected individual pedestrians

Regression-based methods

• Extracting effective features from crowd images
• Utilizing various regression functions to estimate crowd counts
• Edge features, texture features

Perspective information - person scale change along with the perspective geometry

• Blue in the heatmaps indicates small perspective values while yellow indicates large values

High-Level Network Notes

• Generate the GT perspective maps
• Using the K-NN distance to approximate the pedestrian head size
• VGG net backbone
• Three density maps

Paper #2 - Counting Crowded Moving Objects
Key Notes

• Leverage KLT tracker

KLT tracker

• Determine the motion parameters (e.g., affine or pure translation) of local windows W from an image I to a consecutive image J
• KLT runs until no more initial features can be tracked
• Parameter - size of the window

Tracking Challenges

• Inter-object occlusion, self-occlusion, exit from the picture
• Features are agglomeratively clustered into independent objects

Paper #3 - Cross-scene Crowd Counting via Deep Convolutional Neural Networks
Key Notes

• Develop effective features to describe crowd
• Different scenes have different perspective distortions, crowd distributions and lighting conditions.
• CNN Model to detect crowds
• Find Clusters of People
• Apply models to count people in each cluster
• Patches and Density Detection

Paper #4 - Comparison of Tracking Techniques on 360-Degree Videos
Evaluation Criteria

• Viewpoint
• Occlusion
• Deformation
• Lighting
• Scale
• Shakiness

Summary of Trackers

Paper #5 - Beyond Counting: Comparisons of Density Maps for Crowd Analysis Tasks - Counting, Detection, and Tracking
Key Notes

• Counting 
• Clustering of points
• Creation of Density Map

Keep Thinking!!!