Changelog¶

v1.0.0rc1 (10/10/2022)¶

MMTracking 1.0.0rc1 is the 2-nd version of MMTracking 1.x, a part of the OpenMMLab 2.0 projects.

Built upon the new training engine, MMTracking 1.x unifies the interfaces of datasets, models, evaluation, and visualization.

And there are some BC-breaking changes. Please check the migration tutorial for more details.

We also support more methods in MMTracking 1.x, such as StrongSORT for MOT, Mask2Former for VIS, PrDiMP for SOT.

v0.13.0 (29/04/2022)¶

Highlights¶

Support tracking colab tutorial (#511)

New Features¶

Refactor the training datasets of SiamRPN++ (#496), (#518)
Support loading data from ceph for SOT datasets (#494)
Support loading data from ceph for MOT challenge dataset (#517)
Support evaluation metric for VIS task (#501)

Bug Fixes¶

Fix a bug in the LaSOT datasets and update the pretrained models of STARK (#483), (#503)
Fix a bug in the format_results function of VIS task (#504)

v0.12.0 (01/04/2022)¶

Highlights¶

Support QDTrack algorithm in MOT (#433), (#451), (#461), (#469)

Bug Fixes¶

Support empty tensor for selsa aggregator (#463)

v0.11.0 (04/03/2022)¶

Highlights¶

Support STARK algorithm in SOT (#443), (#440), (#434), (#438), (#435), (#426)
Support HOTA evaluation metrics for MOT (#417)

New Features¶

Support TAO dataset in MOT (#415)

v0.10.0 (10/02/2022)¶

New Features¶

Support CPU training (#404)

Improvements¶

Refactor SOT datasets (#401), (#402), (#393)

v0.9.0 (05/01/2022)¶

Highlights¶

Support arXiv 2021 manuscript ‘ByteTrack: Multi-Object Tracking by Associating Every Detection Box’ (#385), (#383), (#372)
Support ICCV 2019 paper ‘Video Instance Segmentation’ (#304), (#303), (#298), (#292)

New Features¶

Support CrowdHuman dataset for MOT (#366)
Support VOT2018 dataset for SOT (#305)
Support YouTube-VIS dataset for VIS (#290)

Bug Fixes¶

Fix two significant bugs in SOT and provide new SOT pretrained models (#349)

Improvements¶

Refactor LaSOT, TrackingNet dataset and support GOT-10K datasets (#296)
Support persisitent workers (#348)

v0.8.0 (03/10/2021)¶

New Features¶

Support OTB100 dataset in SOT (#271)
Support TrackingNet dataset in SOT (#268)
Support UAV123 dataset in SOT (#260)

Bug Fixes¶

Fix a bug in mot_param_search.py (#270)

Improvements¶

Use PyTorch sphinx theme (#274)
Use pycocotools instead of mmpycocotools (#263)

v0.7.0 (03/09/2021)¶

Highlights¶

Release code of AAAI 2021 paper ‘Temporal ROI Align for Video Object Recognition’ (#247)
Refactor English documentations (#243)
Add Chinese documentations (#248), (#250)

New Features¶

Support fp16 training and testing (#230)
Release model using ResNeXt-101 as backbone for all VID methods (#254)
Support the results of Tracktor on MOT15, MOT16 and MOT20 datasets (#217)
Support visualization for single gpu test (#216)

Bug Fixes¶

Fix a bug in MOTP evaluation (#235)
Fix two bugs in reid training and testing (#249)

Improvements¶

Refactor anchor in SiameseRPN++ (#229)
Unify model initialization (#235)
Refactor unittest (#231)

v0.6.0 (30/07/2021)¶

Highlights¶

Fix training bugs of all three tasks (#219), (#221)

New Features¶

Support error visualization for mot task (#212)

Bug Fixes¶

Fix a bug in SOT demo (#213)

Improvements¶

Use MMCV registry (#220)
Add README.md for reid training (#210)
Modify dict keys of the outputs of SOT (#223)
Add Chinese docs including install.md, quick_run.md, model_zoo.md, dataset.md (#205), (#214)

v0.5.3 (01/07/2021)¶

New Features¶

Support ReID training (#177), (#179), (#180), (#181),
Support MIM (#158)

Bug Fixes¶

Fix evaluation hook (#176)
Fix a typo in vid config (#171)

Improvements¶

Refactor nms config (#167)

v0.5.2 (03/06/2021)¶

Improvements¶

Fixed typos (#104, #121, #145)
Added conference reference (#111)
Updated the link of CONTRIBUTING to mmcv (#112)
Adapt updates in mmcv (FP16Hook) (#114, #119)
Added bibtex and links to other codebases (#122)
Added docker files (#124)
Used collect_env in mmcv (#129)
Added and updated Chinese README (#135, #147, #148)

v0.5.1 (01/02/2021)¶

Bug Fixes¶

Fixed ReID checkpoint loading (#80)
Fixed empty tensor in track_result (#86)
Fixed wait_time in MOT demo script (#92)

Improvements¶

Support single-stage detector for DeepSORT (#100)

v0.5.0 (04/01/2021)¶

Highlights¶

MMTracking is released!

New Features¶

Support video object detection methods: DFF, FGFA, SELSA
Support multi object tracking methods: SORT/DeepSORT, Tracktor
Support single object tracking methods: SiameseRPN++