TrackFormer: Multi-Object Tracking with Transformers

preview_player
Показать описание
Following DETR's approach for object detection using transformers, TrackFormer employs them for multi-object tracking given an input video. This has the potential to be beneficial in surveillance, autonomous vehicles, sports analysis, and other applications. In this YouTube video, I present how DETR detects objects and how TrackFormer extends the idea for the task of multi-object tracking.

Table of Content:
00:00 Introduction
00:15 Previous Attempts
02:17 DETR
08:07 TrackFormer
13:47 Bipartite Matching
21:51 Set Prediction Loss
23:31 Track Augmentation
26:39 Result

  1. Soroush Mehraban
  2. TrackFormer
  3. Object tracking
  4. Object Detection
  5. DETR
  6. Transformers
Рекомендации по теме
TrackFormer: Multi-Object Tracking 0:28:40

TrackFormer: Multi-Object Tracking with Transformers

Trackformer - Multi-Object 0:46:31

Trackformer - Multi-Object Tracking with Transformers

[WACV 2024] Contrastive 0:07:59

[WACV 2024] Contrastive Learning for Multi-Object Tracking With Transformers

Trackformer Explained 1:17:08

Trackformer Explained

Contrastive Learning for 0:07:59

Contrastive Learning for Multi-Object Tracking With Transformers

MOTR - Object 0:48:20

MOTR - Object Tracking with Transformers

InterTrack: Interaction Transformer 0:01:16

InterTrack: Interaction Transformer for 3D Multi-Object Tracking

Reading Trackformer source 1:41:54

Reading Trackformer source code

MeMOT: Multi Object 0:04:54

MeMOT: Multi Object Tracking With Memory | CVPR 2022

TransMOT: Spatial-Temporal Graph 0:03:22

TransMOT: Spatial-Temporal Graph Transformer for Multiple Object Tracking

Efficient Visual Tracking 0:03:41

Efficient Visual Tracking with Exemplar Transformers

[CVPR 2023] Generalized 0:04:00

[CVPR 2023] Generalized Relation Modeling for Transformer Tracking

HW3 TrackFormer MOT 0:00:34

HW3 TrackFormer MOT with cv UI Demo

Detecting objects with 0:01:01

Detecting objects with DEtection TRansformer (DETR)

AttTrack: Online Deep 0:04:00

AttTrack: Online Deep Attention Transfer for Multi-object Tracking

[CVPR2023] MOTRv2: Bootstrapping 0:06:48

[CVPR2023] MOTRv2: Bootstrapping End-to-End Multi-Object Tracking by Pretrained Object Detectors

Object Detection Part 0:04:28

Object Detection Part 7: Detection Transformers (DETR), Object Queries

Multiple Object Tracking 0:07:40

Multiple Object Tracking from appearance by hierarchically clustering tracklets - BMVC2022

From Handcrafted to 1:00:26

From Handcrafted to End-to-End Learning, and Back: a Journey for Multi-Object Tracking

Mono-Camera 3D Multi-Object 0:00:11

Mono-Camera 3D Multi-Object Tracking: 0014

Higher Order Multiple 0:01:04

Higher Order Multiple Object Tracking for Crowded Scenes (RVSU 2021)

Object detection Using 0:18:21

Object detection Using Detection Transformer (Detr) on custom dataset

End to End 0:04:59

End to End Referring Video Object Segmentation With Multimodal Transformers | CVPR 2022

Real-Time Object Detection 0:25:41

Real-Time Object Detection Class with Transformers and DETR on Webcam

Комментарии
Автор

This is pure gold! Huge! Thanks for share. invaluable to a project I am working on

HLAI
Автор

the index and the sigma discussion at 19:00at minimum cost mapping is confusing to me, I'll have to rewatch it later

samanthaqiu
Автор

The loss functions were definitely a bit tricky to get around. But that was a really cool video tho!
One thing you could've also touched upon is the usage of deformable detr in place of detr. I can see the trackformer code does incorporate it but wanted to know what changes in trackformer when you switch from detr to deformable one?

subramanyabhat
Автор

Hi, I am currently reading this paper, and I don't know where do the sigma_NMS and the other sigma threshold come from since the output according to this paper is just bounding box and classification result.

Where do we get those sigma value from the output of TrackFormer ?
Or the sigma value is the front 3 value of the classification score vector ?

蔡嘉祥-qm