{ "id": 159392, "title": "BASELINE MODEL", "url": "/competitions/open-images-object-detection-rvc-2020/discussion/159392", "author_name": "littletomatodonkey ", "post_date": "2020-06-17T09:58:49.463000", "votes": 2, "comment_count": 0, "views": 0, "content": "\n\n

After the models' ensemble, the final score on public leaderboard comes the 0.6816. Here all the backbones and pretrained weights come from PaddleClas, which is an image classification toolset.

\n\n

\"\"

\n\n

More details can be seen on PaddleDetection: https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md\nYou can refer to configuration and model link to directly achieve a very high baseline.

\n\n", "messages": [ { "id": 890095, "postDate": "2020-06-17T09:58:49.463Z", "content": "\n\n

After the models' ensemble, the final score on public leaderboard comes the 0.6816. Here all the backbones and pretrained weights come from PaddleClas, which is an image classification toolset.

\n\n

\"\"

\n\n

More details can be seen on PaddleDetection: https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md\nYou can refer to configuration and model link to directly achieve a very high baseline.

\n\n", "rawMarkdown": "* **PaddleDetection** open sourced the best single model last year in Open Images V5, the score is 0.627(single scale), ***0.649***(multi-scale), respectively on Public Leaderboard.\nThe report can be seen here. [https://arxiv.org/pdf/1911.07171.pdf](https://arxiv.org/pdf/1911.07171.pdf)\n\nAfter the models' ensemble, the final score on public leaderboard comes the 0.6816. Here all the backbones and pretrained weights come from [PaddleClas](https://github.com/PaddlePaddle/PaddleClas), which is an image classification toolset.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F997125%2F5a3d333f6fe25049d103295df563c6a4%2Fensemble.png?generation=1592387502929047&alt=media)\n\n\nMore details can be seen on PaddleDetection: [https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md](https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md)\nYou can refer to [configuration](https://github.com/PaddlePaddle/PaddleDetection/blob/master/configs/oidv5/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.yml) and [model link](https://paddlemodels.bj.bcebos.com/object_detection/oidv5_cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.tar) to directly achieve a very high baseline.\n\n* Of course, various of backbones are alse vital for the final model ensemble.\n **PaddleClas** open sourced 117 pretrained models on ImageNet1k dataset, which are important for the competition, you can refer to [https://github.com/PaddlePaddle/PaddleClas](https://github.com/PaddlePaddle/PaddleClas) to easily use the pretrained models and help to improve the detection models' performance. The models are avaiable here: [https://github.com/PaddlePaddle/PaddleClas/blob/master/docs/en/models/models_intro.md](https://github.com/PaddlePaddle/PaddleClas/blob/master/docs/en/models/models_intro.md)\n\n* Last but not least, PaddleClas open sourced a **83.7%** Top-1 Acc [ResNet101_vd pretrained model](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_ssld_pretrained.tar), which is proved to perfom well combining with CBNet. So welcome to use it and get high scores~\n\n", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "890095": "* **PaddleDetection** open sourced the best single model last year in Open Images V5, the score is 0.627(single scale), ***0.649***(multi-scale), respectively on Public Leaderboard.\nThe report can be seen here. [https://arxiv.org/pdf/1911.07171.pdf](https://arxiv.org/pdf/1911.07171.pdf)\n\nAfter the models' ensemble, the final score on public leaderboard comes the 0.6816. Here all the backbones and pretrained weights come from [PaddleClas](https://github.com/PaddlePaddle/PaddleClas), which is an image classification toolset.\n\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F997125%2F5a3d333f6fe25049d103295df563c6a4%2Fensemble.png?generation=1592387502929047&alt=media)\n\n\nMore details can be seen on PaddleDetection: [https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md](https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/champion_model/OIDV5_BASELINE_MODEL.md)\nYou can refer to [configuration](https://github.com/PaddlePaddle/PaddleDetection/blob/master/configs/oidv5/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.yml) and [model link](https://paddlemodels.bj.bcebos.com/object_detection/oidv5_cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.tar) to directly achieve a very high baseline.\n\n* Of course, various of backbones are alse vital for the final model ensemble.\n **PaddleClas** open sourced 117 pretrained models on ImageNet1k dataset, which are important for the competition, you can refer to [https://github.com/PaddlePaddle/PaddleClas](https://github.com/PaddlePaddle/PaddleClas) to easily use the pretrained models and help to improve the detection models' performance. The models are avaiable here: [https://github.com/PaddlePaddle/PaddleClas/blob/master/docs/en/models/models_intro.md](https://github.com/PaddlePaddle/PaddleClas/blob/master/docs/en/models/models_intro.md)\n\n* Last but not least, PaddleClas open sourced a **83.7%** Top-1 Acc [ResNet101_vd pretrained model](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_ssld_pretrained.tar), which is proved to perfom well combining with CBNet. So welcome to use it and get high scores~\n\n" } }