MMSegmentation v0.20.0 Release

Highlights

• Support Twins (#989)
• Support a real-time segmentation model STDC (#995)
• Support a widely-used segmentation model in lane detection ERFNet (#960)
• Support A Remote Sensing Land-Cover Dataset LoveDA (#1028)
• Support focal loss (#1024)

New Features

• Support Twins (#989)
• Support a real-time segmentation model STDC (#995)
• Support a widely-used segmentation model in lane detection ERFNet (#960)
• Add SETR cityscapes benchmark (#1087)
• Add BiSeNetV1 COCO-Stuff 164k benchmark (#1019)
• Support focal loss (#1024)
• Add Cutout transform (#1022)

Improvements

• Set a random seed when the user does not set a seed (#1039)
• Add CircleCI setup (#1086)
• Skip CI on ignoring given paths (#1078)
• Add abstract and image for every paper (#1060)
• Create a symbolic link on windows (#1090)
• Support video demo using trained model (#1014)

Bug Fixes

• Fix incorrectly loading init_cfg or pretrained models of several transformer models (#999, #1069, #1102)
• Fix EfficientMultiheadAttention in SegFormer (#1003)
• Remove fp16 folder in configs (#1031)
• Fix several typos in .yml file (Dice Metric #1041, ADE20K dataset #1120, Training Memory (GB) #1083)
• Fix test error when using --show-dir (#1091)
• Fix dist training infinite waiting issue (#1035)
• Change the upper version of mmcv to 1.5.0 (#1096)
• Fix symlink failure on Windows (#1038)
• Cancel previous runs that are not completed (#1118)
• Unified links of readthedocs in docs (#1119)

Contributors

• @Junjue-Wang made their first contribution in #1028
• @ddebby made their first contribution in #1066
• @del-zhenwu made their first contribution in #1078
• @KangBK0120 made their first contribution in #1106
• @zergzzlun made their first contribution in #1091
• @fingertap made their first contribution in #1035
• @irvingzhang0512 made their first contribution in #1014
• @littleSunlxy made their first contribution in #989
• @lkm2835
• @RockeyCoss
• @MengzhangLI
• @Junjun2016
• @xiexinch
• @xvjiarui

MMSegmentation v0.19.0 Release

Highlights

• Support TIMMBackbone wrapper (#998)
• Support custom hook (#428)
• Add codespell pre-commit hook (#920)
• Add FastFCN benchmark on ADE20K (#972)

New Features

• Support TIMMBackbone wrapper (#998)
• Support custom hook (#428)
• Add FastFCN benchmark on ADE20K (#972)
• Add codespell pre-commit hook and fix typos (#920)

Improvements

• Make inputs & channels smaller in unittests (#1004)
• Change self.loss_decode back to dict in Single Loss situation (#1002)

Bug Fixes

• Fix typo in usage example (#1003)
• Add contiguous after permutation in ViT (#992)
• Fix the invalid link (#985)
• Fix bug in CI with python 3.9 (#994)
• Fix bug when loading class name form file in custom dataset (#923)

Contributors

• @ShoupingShan made their first contribution in #923
• @RockeyCoss made their first contribution in #954
• @HarborYuan made their first contribution in #992
• @lkm2835 made their first contribution in #1003
• @gszh made their first contribution in #428
• @xvjiarui
• @VVsssssk
• @MengzhangLI
• @Junjun2016

MMSegmentation v0.18.0 Release

Highlights

• Support three real-time segmentation models (ICNet #884, BiSeNetV1 #851, and BiSeNetV2 #804)
• Support one efficient segmentation model (FastFCN #885)
• Support one efficient non-local/self-attention based segmentation model (ISANet #70)
• Support COCO-Stuff 10k and 164k datasets (#625)
• Support evaluate concated dataset separately (#833)
• Support loading GT for evaluation from multi-file backend (#867)

New Features

• Support three real-time segmentation models (ICNet #884, BiSeNetV1 #851, and BiSeNetV2 #804)
• Support one efficient segmentation model (FastFCN #885)
• Support one efficient non-local/self-attention based segmentation model (ISANet #70)
• Support COCO-Stuff 10k and 164k datasets (#625)
• Support evaluate concated dataset separately (#833)

Improvements

• Support loading GT for evaluation from multi-file backend (#867)
• Auto-convert SyncBN to BN when training on DP automatly(#772)
• Refactor Swin-Transformer (#800)

Bug Fixes

• Update mmcv installation in dockerfile (#860)
• Fix number of iteration bug when resuming checkpoint in distributed train (#866)
• Fix parsing parse in val_step (#906)

MMSegmentation v0.17.0 Release

Highlights

• Support SegFormer
• Support DPT
• Support Dark Zurich and Nighttime Driving datasets
• Support progressive evaluation

New Features

• Support SegFormer (#599)
• Support DPT (#605)
• Support Dark Zurich and Nighttime Driving datasets (#815)
• Support progressive evaluation (#709)

Improvements

• Add multiscale_output interface and unittests for HRNet (#830)
• Support inherit cityscapes dataset (#750)
• Fix some typos in README.md (#824)
• Delete convert function and add instruction to ViT/Swin README.md (#791)
• Add vit/swin/mit convert weight scripts (#783)
• Add copyright files (#796)

Bug Fixes

• Fix invalid checkpoint link in inference_demo.ipynb (#814)
• Ensure that items in dataset have the same order across multi machine (#780)
• Fix the log error (#766)

MMSegmentation v0.16.0 Release

Highlights

• Support PyTorch 1.9
• Support SegFormer backbone MiT
• Support md2yml pre-commit hook
• Support frozen stage for HRNet

New Features

• Support SegFormer backbone MiT (#594)
• Support md2yml pre-commit hook (#732)
• Support mim (#717)
• Add mmseg2torchserve tool (#552)

Improvements

• Support hrnet frozen stage (#743)
• Add template of reimplementation questions (#741)
• Output pdf and epub formats for readthedocs (#742)
• Refine the docstring of ResNet (#723)
• Replace interpolate with resize (#731)
• Update resource limit (#700)
• Update config.md (#678)

Bug Fixes

• Fix ATTENTION registry (#729)
• Fix analyze log script (#716)
• Fix doc api display (#725)
• Fix patch_embed and pos_embed mismatch error (#685)
• Fix efficient test for multi-node (#707)
• Fix init_cfg in resnet backbone (#697)
• Fix efficient test bug (#702)
• Fix url error in config docs (#680)
• Fix mmcv installation (#676)
• Fix torch version (#670)

Contributors

@sshuair @xiexinch @Junjun2016 @mmeendez8 @xvjiarui @Sennnnn @puhsu @BIGWangYuDong @keke1u @daavoo

MMSegmentation v0.15.0 Release

Highlights

• Support ViT, SETR, and Swin-Transformer
• Add Chinese documentation
• Unified parameter initialization

Bug Fixes

• Fix typo and links (#608)
• Fix Dockerfile (#607)
• Fix ViT init (#609)
• Fix mmcv version compatible table (#658)
• Fix model links of DMNet and UNet (#660)

New Features

• Support loading DeiT weights (#538)
• Support SETR (#531, #635)
• Add config and models for ViT backbone with UperHead (#520, #635)
• Support Swin-Transformer (#511)
• Add higher accuracy FastSCNN (#606)
• Add Chinese documentation (#666)

Improvements

• Unified parameter initialization (#567)
• Separate CUDA and CPU in github action CI (#602)
• Support persistent dataloader worker (#646)
• Update meta file fields (#661, #664)

MMSegmentation v0.14.0 Release

Highlights

• Support ONNX to TensorRT
• Support MIM

Bug Fixes

• Fix ONNX to TensorRT verify (#547)
• Fix save best for EvalHook (#575)

New Features

• Support loading DeiT weights (#538)
• Support ONNX to TensorRT (#542)
• Support output results for ADE20k (#544)
• Support MIM (#549)

Improvements

• Add option for ViT output shape (#530)
• Infer batch size using len(result) (#532)
• Add compatible table between MMSeg and MMCV (#558)

MMSegmentation v0.13.0 Release

Highlights

• Support Pascal Context Class-59 dataset.
• Support Visual Transformer Backbone.
• Support mFscore metric.

Bug Fixes

• Fixed Colaboratory tutorial (#451)
• Fixed mIoU calculation range (#471)
• Fixed sem_fpn, unet README.md (#492)
• Fixed num_classes in FCN for Pascal Context 60-class dataset (#488)
• Fixed FP16 inference (#497)

New Features

• Support dynamic export and visualize to pytorch2onnx (#463)
• Support export to torchscript (#469, #499)
• Support Pascal Context Class-59 dataset (#459)
• Support Visual Transformer backbone (#465)
• Support UpSample Neck (#512)
• Support mFscore metric (#509)

Improvements

• Add more CI for PyTorch (#460)
• Add print model graph args for tools/print_config.py (#451)
• Add cfg links in modelzoo README.md (#468)
• Add BaseSegmentor import to segmentors/init.py (#495)
• Add MMOCR, MMGeneration links (#501, #506)
• Add Chinese QR code (#506)
• Use MMCV MODEL_REGISTRY (#515)
• Add ONNX testing tools (#498)
• Replace data_dict calling 'img' key to support MMDet3D (#514)
• Support reading class_weight from file in loss function (#513)
• Make tags as comment (#505)
• Use MMCV EvalHook (#438)

MMSegmentation v0.12.0 Release

Highlights

• Support FCN-Dilate 6 model.
• Support Dice Loss.

Bug Fixes

• Fixed PhotoMetricDistortion Doc (#388)
• Fixed install scripts (#399)
• Fixed Dice Loss multi-class (#417)

New Features

• Support Dice Loss (#396)
• Add plot logs tool (#426)
• Add opacity option to show_result (#425)
• Speed up mIoU metric (#430)

Improvements

• Refactor unittest file structure (#440)
• Fix typos in the repo (#449)
• Include class-level metrics in the log (#445)

MMSegmentation v0.11.0 Release

Highlights

• Support memory efficient test, add more UNet models.

Bug Fixes

• Fixed TTA resize scale (#334)
• Fixed CI for pip 20.3 (#307)
• Fixed ADE20k test (#359)

New Features

• Support memory efficient test (#330)
• Add more UNet benchmarks (#324)
• Support Lovasz Loss (#351)

Improvements

• Move train_cfg/test_cfg inside model (#341)