"No one is harder on a talented person than the person themselves" - Linda Wilkinson ; "Trust your guts and don't follow the herd" ; "Validate direction not destination" ;

January 10, 2021

Tracking @ 2021 - Paper Reads - Object Tracking

Paper - Object Detection and Tracking Algorithms for Vehicle Counting: A Comparative Analysis

Key Notes

  • Models used - CenterNet and Deep SORT, Detectron2 and Deep
  • SORT, and YOLOv4 and Deep SORT
  • YOLO and SORT algorithms
  • Counting Techniques
  • Counting by frame differencing
  • Counting by detection
  • Motion based counting
  • Deep learning based counting

OBJECT DETECTORS

CenterNet

  • CenterNet functions on the intuition that if a detected bounding box has a higher Intersection over Union (IoU) with the ground-truth box
  • CenterNet is a singlestage detector 

Detectron2

Detectron2 supports implementation to multiple object detection algorithms using different backbone network architectures such as ResNET {50, 101, 152}, FPN, VGG16

YOLOv4

You Only Look Once (YOLO) is the state-of-the-art object detection algorithm. New techniques adopted in YOLOv4 are: (i) WeightedResidual-Connections, (ii) Cross-Stage-Partial-Connections, (iii) Cross mini-batch, (iv) Normalization (CmBN), (v) Selfadversial-training, (vi) Mish-activation

EfficientDet

Follows single-stage detectors pattern

SORT

Simple Online and Realtime Tracking (SORT) is an implementation of tracking-by-detection framework where the main objective is to detect objects each frame and associate them for online and real-time tracking application

Deep SORT

A combination of Kalman Filter and Hungarian algorithm is used for tracking

Paper - Tracking Objects as Points

Key Notes

  • We track objects by tracking their centers. We learn a 2D offset between two adjacent frames and associate them based on center distance.
  • A simple displacement prediction, akin to sparse optical flow, allows objects in different frames to be linked
  • Joint detection and tracking
  • Early approaches [2, 47] used Kalman filters to model object velocities

Paper - SMOT: Single-Shot Multi Object Tracking

Key Notes

  • The first stage exercises a per-frame object detector to localize object bounding boxes in each frame
  • The second stage, tracklet generation, merges detection results to create a set of tracklets, i.e., short tracks, based on short-term cues

More Read

FairMOT: On the Fairness of Detection and Re-Identification in Multiple Object Tracking

Object Detection and Tracking in 2020


Paper - ByteTrack: Multi-Object Tracking

Key Notes

In this paper, we identify that the similarity with tracklets provides a strong cue to distinguish the objects and background in low score detection boxes

Data association is the core of multi-object tracking, which first computes the similarity between tracklets and detection boxes and then matches them according to the similarity.

Location and motion similarity are accurate in the short-range matching.

Code

Codes

Happy Learning!!!

No comments: