# Model Zoo and Baselines # Introduction We provide baselines trained with Detectron2 DensePose. The corresponding configuration files can be found in the [configs](../configs) directory. All models were trained on COCO `train2014` + `valminusminival2014` and evaluated on COCO `minival2014`. For the details on common settings in which baselines were trained, please check [Detectron 2 Model Zoo](../../../MODEL_ZOO.md). ## License All models available for download through this document are licensed under the [Creative Commons Attribution-ShareAlike 3.0 license](https://creativecommons.org/licenses/by-sa/3.0/) ## COCO DensePose Baselines with DensePose-RCNN ### Legacy Models Baselines trained using schedules from [Güler et al, 2018](https://arxiv.org/pdf/1802.00434.pdf)
Name lr
sched		 train
time
(s/iter)		 inference
time
(s/im)		 train
mem
(GB)		 box
AP		 dp. AP
GPS		 dp. AP
GPSm		 model id download
R_50_FPN_s1x_legacy s1x 0.307 0.051 3.2 58.1 52.1 54.9 164832157 model | metrics
R_101_FPN_s1x_legacy s1x 0.390 0.063 4.3 59.5 53.2 56.1 164832182 model | metrics
### Improved Baselines, Original Fully Convolutional Haad These models use an improved training schedule and Panoptic FPN head from [Kirillov et al, 2019](https://arxiv.org/abs/1901.02446).
Name lr
sched		 train
time
(s/iter)		 inference
time
(s/im)		 train
mem
(GB)		 box
AP		 dp. AP
GPS		 dp. AP
GPSm		 model id download
R_50_FPN_s1x s1x 0.359 0.066 4.5 61.2 63.7 65.3 165712039 model | metrics
R_101_FPN_s1x s1x 0.428 0.079 5.8 62.3 64.5 66.4 165712084 model | metrics
### Improved Baselines, DeepLabV3 Head These models use an improved training schedule, Panoptic FPN head from [Kirillov et al, 2019](https://arxiv.org/abs/1901.02446) and DeepLabV3 head from [Chen et al, 2017](https://arxiv.org/abs/1706.05587).
Name lr
sched		 train
time
(s/iter)		 inference
time
(s/im)		 train
mem
(GB)		 box
AP		 dp. AP
GPS		 dp. AP
GPSm		 model id download
R_50_FPN_DL_s1x s1x 0.392 0.070 6.7 61.1 65.6 66.8 165712097 model | metrics
R_101_FPN_DL_s1x s1x 0.478 0.083 7.0 62.3 66.3 67.7 165712116 model | metrics
### Baselines with Confidence Estimation These models perform additional estimation of confidence in regressed UV coodrinates, along the lines of [Neverova et al., 2019](https://papers.nips.cc/paper/8378-correlated-uncertainty-for-learning-dense-correspondences-from-noisy-labels).
Name lr
sched		 train
time
(s/iter)		 inference
time
(s/im)		 train
mem
(GB)		 box
AP		 dp. AP
GPS		 dp. AP
GPSm		 model id download
R_50_FPN_WC1_s1x s1x 0.353 0.064 4.6 60.5 64.2 65.6 173862049 model | metrics
R_50_FPN_WC2_s1x s1x 0.364 0.066 4.8 60.7 64.2 65.7 173861455 model | metrics
R_50_FPN_DL_WC1_s1x s1x 0.397 0.068 6.7 61.1 65.8 67.1 173067973 model | metrics
R_50_FPN_DL_WC2_s1x s1x 0.410 0.070 6.8 60.8 65.6 66.7 173859335 model | metrics
R_101_FPN_WC1_s1x s1x 0.435 0.076 5.7 62.5 64.9 66.5 171402969 model | metrics
R_101_FPN_WC2_s1x s1x 0.450 0.078 5.7 62.3 64.8 66.6 173860702 model | metrics
R_101_FPN_DL_WC1_s1x s1x 0.479 0.081 7.9 62.0 66.2 67.4 173858525 model | metrics
R_101_FPN_DL_WC2_s1x s1x 0.491 0.082 7.6 61.7 65.9 67.3 173294801 model | metrics
## Old Baselines It is still possible to use some baselines from [DensePose 1](https://github.com/facebookresearch/DensePose). Below are evaluation metrics for the baselines recomputed in the current framework: | Model | bbox AP | AP | AP50 | AP75 | APm |APl | |-----|-----|-----|--- |--- |--- |--- | | [`ResNet50_FPN_s1x-e2e`](https://dl.fbaipublicfiles.com/densepose/DensePose_ResNet50_FPN_s1x-e2e.pkl) | 54.673 | 48.894 | 84.963 | 50.717 | 43.132 | 50.433 | | [`ResNet101_FPN_s1x-e2e`](https://dl.fbaipublicfiles.com/densepose/DensePose_ResNet101_FPN_s1x-e2e.pkl) | 56.032 | 51.088 | 86.250 | 55.057 | 46.542 | 52.563 | Note: these scores are close, but not strictly equal to the ones reported in the [DensePose 1 Model Zoo](https://github.com/facebookresearch/DensePose/blob/master/MODEL_ZOO.md), which is due to small incompatibilities between the frameworks.