MODEL_ZOO.md 35 KB
English | 简体中文
Model Zoo and Benchmark
Environment
- Python 2.7.1
- PaddlePaddle >=1.5
- CUDA 9.0
- cuDNN >=7.4
- NCCL 2.1.2
Common settings
- All models below were trained on
coco_2017_train, and tested oncoco_2017_val. - Batch Normalization layers in backbones are replaced by Affine Channel layers.
- Unless otherwise noted, all ResNet backbones adopt the ResNet-B variant..
- For RCNN and RetinaNet models, only horizontal flipping data augmentation was used in the training phase and no augmentations were used in the testing phase.
- Inf time (fps): the inference time is measured with fps (image/s) on a single GPU (Tesla V100) with cuDNN 7.5 by running 'tools/eval.py' on all validation set, which including data loadding, network forward and post processing. The batch size is 1.
Training Schedules
- We adopt exactly the same training schedules as Detectron.
- 1x indicates the schedule starts at a LR of 0.02 and is decreased by a factor of 10 after 60k and 80k iterations and eventually terminates at 90k iterations for minibatch size 16. For batch size 8, LR is decreased to 0.01, total training iterations are doubled, and the decay milestones are scaled by 2.
- 2x schedule is twice as long as 1x, with the LR milestones scaled accordingly.
ImageNet Pretrained Models
The backbone models pretrained on ImageNet are available. All backbone models are pretrained on standard ImageNet-1k dataset and can be downloaded here.
- Notes: The ResNet50 model was trained with cosine LR decay schedule and can be downloaded here.
Baselines
Faster & Mask R-CNN
| Backbone | Type | Image/gpu | Lr schd | Inf time (fps) | Box AP | Mask AP | Download | Configs |
|---|---|---|---|---|---|---|---|---|
| ResNet50 | Faster | 1 | 1x | 12.747 | 35.2 | - | model | config |
| ResNet50 | Faster | 1 | 2x | 12.686 | 37.1 | - | model | config |
| ResNet50 | Mask | 1 | 1x | 11.615 | 36.5 | 32.2 | model | config |
| ResNet50 | Mask | 1 | 2x | 11.494 | 38.2 | 33.4 | model | config |
| ResNet50-vd | Faster | 1 | 1x | 12.575 | 36.4 | - | model | config |
| ResNet34-FPN | Faster | 2 | 1x | - | 36.7 | - | model | config |
| ResNet34-vd-FPN | Faster | 2 | 1x | - | 37.4 | - | model | config |
| ResNet50-FPN | Faster | 2 | 1x | 22.273 | 37.2 | - | model | config |
| ResNet50-FPN | Faster | 2 | 2x | 22.297 | 37.7 | - | model | config |
| ResNet50-FPN | Mask | 1 | 1x | 15.184 | 37.9 | 34.2 | model | config |
| ResNet50-FPN | Mask | 1 | 2x | 15.881 | 38.7 | 34.7 | model | config |
| ResNet50-FPN | Cascade Faster | 2 | 1x | 17.507 | 40.9 | - | model | config |
| ResNet50-FPN | Cascade Mask | 1 | 1x | 12.43 | 41.3 | 35.5 | model | config |
| ResNet50-vd-FPN | Faster | 2 | 2x | 21.847 | 38.9 | - | model | config |
| ResNet50-vd-FPN | Mask | 1 | 2x | 15.825 | 39.8 | 35.4 | model | config |
| CBResNet50-vd-FPN | Faster | 2 | 1x | - | 39.7 | - | model | config |
| ResNet101 | Faster | 1 | 1x | 9.316 | 38.3 | - | model | config |
| ResNet101-FPN | Faster | 1 | 1x | 17.297 | 38.7 | - | model | config |
| ResNet101-FPN | Faster | 1 | 2x | 17.246 | 39.1 | - | model | config |
| ResNet101-FPN | Mask | 1 | 1x | 12.983 | 39.5 | 35.2 | model | config |
| ResNet101-vd-FPN | Faster | 1 | 1x | 17.011 | 40.5 | - | model | config |
| ResNet101-vd-FPN | Faster | 1 | 2x | 16.934 | 40.8 | - | model | config |
| ResNet101-vd-FPN | Mask | 1 | 1x | 13.105 | 41.4 | 36.8 | model | config |
| CBResNet101-vd-FPN | Faster | 2 | 1x | - | 42.7 | - | model | config |
| ResNeXt101-vd-64x4d-FPN | Faster | 1 | 1x | 8.815 | 42.2 | - | model | config |
| ResNeXt101-vd-64x4d-FPN | Faster | 1 | 2x | 8.809 | 41.7 | - | model | config |
| ResNeXt101-vd-64x4d-FPN | Mask | 1 | 1x | 7.689 | 42.9 | 37.9 | model | config |
| ResNeXt101-vd-64x4d-FPN | Mask | 1 | 2x | 7.859 | 42.6 | 37.6 | model | config |
| SENet154-vd-FPN | Faster | 1 | 1.44x | 3.408 | 42.9 | - | model | config |
| SENet154-vd-FPN | Mask | 1 | 1.44x | 3.233 | 44.0 | 38.7 | model | config |
| ResNet101-vd-FPN | CascadeClsAware Faster | 2 | 1x | - | 44.7(softnms) | - | model | config |
| ResNet101-vd-FPN | CascadeClsAware Faster | 2 | 1x | - | 46.5(multi-scale test) | - | model | config |
Deformable ConvNets v2
| Backbone | Type | Conv | Image/gpu | Lr schd | Inf time (fps) | Box AP | Mask AP | Download | Configs |
|---|---|---|---|---|---|---|---|---|---|
| ResNet50-FPN | Faster | c3-c5 | 2 | 1x | 19.978 | 41.0 | - | model | config |
| ResNet50-vd-FPN | Faster | c3-c5 | 2 | 2x | 19.222 | 42.4 | - | model | config |
| ResNet101-vd-FPN | Faster | c3-c5 | 2 | 1x | 14.477 | 44.1 | - | model | config |
| ResNeXt101-vd-64x4d-FPN | Faster | c3-c5 | 1 | 1x | 7.209 | 45.2 | - | model | config |
| ResNet50-FPN | Mask | c3-c5 | 1 | 1x | 14.53 | 41.9 | 37.3 | model | config |
| ResNet50-vd-FPN | Mask | c3-c5 | 1 | 2x | 14.832 | 42.9 | 38.0 | model | config |
| ResNet101-vd-FPN | Mask | c3-c5 | 1 | 1x | 11.546 | 44.6 | 39.2 | model | config |
| ResNeXt101-vd-64x4d-FPN | Mask | c3-c5 | 1 | 1x | 6.45 | 46.2 | 40.4 | model | config |
| ResNet50-FPN | Cascade Faster | c3-c5 | 2 | 1x | - | 44.2 | - | model | config |
| ResNet101-vd-FPN | Cascade Faster | c3-c5 | 2 | 1x | - | 46.4 | - | model | config |
| ResNeXt101-vd-FPN | Cascade Faster | c3-c5 | 2 | 1x | - | 47.3 | - | model | config |
| SENet154-vd-FPN | Cascade Mask | c3-c5 | 1 | 1.44x | - | 51.9 | 43.9 | model | config |
| ResNet200-vd-FPN-Nonlocal | CascadeClsAware Faster | c3-c5 | 1 | 2.5x | 3.103 | 51.7%(softnms) | - | model | config |
| CBResNet200-vd-FPN-Nonlocal | Cascade Faster | c3-c5 | 1 | 2.5x | 1.68 | 53.3%(softnms) | - | model | config |
Notes:
- Deformable ConvNets v2(dcn_v2) reference from Deformable ConvNets v2.
c3-c5means addingdcnin resnet stage 3 to 5.- Detailed configuration file in configs/dcn
HRNet
- See more details in HRNet model zoo.
Res2Net
- See more details in Res2Net model zoo.
IOU loss
- GIOU loss and DIOU loss are included now. See more details in IOU loss model zoo.
GCNet
- See more details in GCNet model zoo.
Libra R-CNN
- See more details in Libra R-CNN model zoo.
Auto Augmentation
- See more details in Auto Augmentation model zoo.
Group Normalization
| Backbone | Type | Image/gpu | Lr schd | Box AP | Mask AP | Download | Configs |
|---|---|---|---|---|---|---|---|
| ResNet50-FPN | Faster | 2 | 2x | 39.7 | - | model | config |
| ResNet50-FPN | Mask | 1 | 2x | 40.1 | 35.8 | model | config |
Notes:
- Group Normalization reference from Group Normalization.
- Detailed configuration file in configs/gn
YOLO v3
| Backbone | Pretrain dataset | Size | deformable Conv | Image/gpu | Lr schd | Inf time (fps) | Box AP | Download | Configs |
|---|---|---|---|---|---|---|---|---|---|
| DarkNet53 (paper) | ImageNet | 608 | False | 8 | 270e | - | 33.0 | - | - |
| DarkNet53 (paper) | ImageNet | 416 | False | 8 | 270e | - | 31.0 | - | - |
| DarkNet53 (paper) | ImageNet | 320 | False | 8 | 270e | - | 28.2 | - | - |
| DarkNet53 | ImageNet | 608 | False | 8 | 270e | 45.571 | 38.9 | model | config |
| DarkNet53 | ImageNet | 416 | False | 8 | 270e | - | 37.5 | model | config |
| DarkNet53 | ImageNet | 320 | False | 8 | 270e | - | 34.8 | model | config |
| MobileNet-V1 | ImageNet | 608 | False | 8 | 270e | 78.302 | 29.3 | model | config |
| MobileNet-V1 | ImageNet | 416 | False | 8 | 270e | - | 29.3 | model | config |
| MobileNet-V1 | ImageNet | 320 | False | 8 | 270e | - | 27.1 | model | config |
| MobileNet-V3 | ImageNet | 608 | False | 8 | 270e | - | 31.6 | model | config |
| MobileNet-V3 | ImageNet | 416 | False | 8 | 270e | - | 29.9 | model | config |
| MobileNet-V3 | ImageNet | 320 | False | 8 | 270e | - | 27.1 | model | config |
| ResNet34 | ImageNet | 608 | False | 8 | 270e | 63.356 | 36.2 | model | config |
| ResNet34 | ImageNet | 416 | False | 8 | 270e | - | 34.3 | model | config |
| ResNet34 | ImageNet | 320 | False | 8 | 270e | - | 31.4 | model | config |
| ResNet50_vd | ImageNet | 608 | True | 8 | 270e | - | 39.1 | model | config |
| ResNet50_vd | Object365 | 608 | True | 8 | 270e | - | 41.4 | model | config |
YOLO v3 on Pascal VOC
| Backbone | Size | Image/gpu | Lr schd | Inf time (fps) | Box AP(0.5) | Download | Configs |
|---|---|---|---|---|---|---|---|
| DarkNet53 | 608 | 8 | 270e | 54.977 | 83.5 | model | config |
| DarkNet53 | 416 | 8 | 270e | - | 83.6 | model | config |
| DarkNet53 | 320 | 8 | 270e | - | 82.2 | model | config |
| DarkNet53 Diou-Loss | 608 | 8 | 270e | - | 83.5 | model | config |
| MobileNet-V1 | 608 | 8 | 270e | 104.291 | 76.2 | model | config |
| MobileNet-V1 | 416 | 8 | 270e | - | 76.7 | model | config |
| MobileNet-V1 | 320 | 8 | 270e | - | 75.3 | model | config |
| ResNet34 | 608 | 8 | 270e | 82.247 | 82.6 | model | config |
| ResNet34 | 416 | 8 | 270e | - | 81.9 | model | config |
| ResNet34 | 320 | 8 | 270e | - | 80.1 | model | config |
Notes:
- YOLOv3-DarkNet53 performance in paper YOLOv3 is also provided above, our implements improved performance mainly by using L1 loss in bounding box width and height regression, image mixup and label smooth.
- YOLO v3 is trained in 8 GPU with total batch size as 64 and trained 270 epoches. YOLO v3 training data augmentations: mixup, randomly color distortion, randomly cropping, randomly expansion, randomly interpolation method, randomly flippling. YOLO v3 used randomly reshaped minibatch in training, inferences can be performed on different image sizes with the same model weights, and we provided evaluation results of image size 608/416/320 above. Deformable conv is added on stage 5 of backbone.
- Compared with YOLOv3-DarkNet53, the average AP of YOLOv3-DarkNet53 with Diou-Loss increases about 2% in VOC dataset.
- YOLO v3 enhanced model improves the precision to 43.6 involved with deformable conv, dropblock, IoU loss and IoU aware. See more details in YOLOv3_ENHANCEMENT
RetinaNet
| Backbone | Image/gpu | Lr schd | Inf time (fps) | Box AP | Download | Configs |
|---|---|---|---|---|---|---|
| ResNet50-FPN | 2 | 1x | - | 36.0 | model | config |
| ResNet101-FPN | 2 | 1x | - | 37.3 | model | config |
| ResNeXt101-vd-FPN | 1 | 1x | - | 40.5 | model | config |
Notes: In RetinaNet, the base LR is changed to 0.01 for minibatch size 16.
EfficientDet
| Scale | Image/gpu | Lr schd | Box AP | Download |
|---|---|---|---|---|
| EfficientDet-D0 | 16 | 300 epochs | 33.8 | model |
Notes: base LR is 0.16 for minibatch size 128 (8x16).
SSDLite
| Backbone | Size | Image/gpu | Lr schd | Inf time (fps) | Box AP | Download | Configs |
|---|---|---|---|---|---|---|---|
| MobileNet_v1 | 300 | 64 | Cosine decay(40w) | - | 23.6 | model | config |
| MobileNet_v3 small | 320 | 64 | Cosine decay(40w) | - | 16.2 | model | config |
| MobileNet_v3 large | 320 | 64 | Cosine decay(40w) | - | 23.3 | model | config |
| MobileNet_v3 small w/ FPN | 320 | 64 | Cosine decay(40w) | - | 18.9 | model | config |
| MobileNet_v3 large w/ FPN | 320 | 64 | Cosine decay(40w) | - | 24.3 | model | config |
| GhostNet | 320 | 64 | Cosine decay(40w) | - | 23.3 | model | config |
Notes: SSDLite is trained in 8 GPU with total batch size as 512 and uses cosine decay strategy to train.
SSD
| Backbone | Size | Image/gpu | Lr schd | Inf time (fps) | Box AP | Download | Configs |
|---|---|---|---|---|---|---|---|
| VGG16 | 300 | 8 | 40w | 81.613 | 25.1 | model | config |
| VGG16 | 512 | 8 | 40w | 46.007 | 29.1 | model | config |
Notes: VGG-SSD is trained in 4 GPU with total batch size as 32 and trained 400000 iters.
SSD on Pascal VOC
| Backbone | Size | Image/gpu | Lr schd | Inf time (fps) | Box AP(0.5) | Download | Configs |
|---|---|---|---|---|---|---|---|
| MobileNet v1 | 300 | 32 | 120e | 159.543 | 73.2 | model | config |
| VGG16 | 300 | 8 | 240e | 117.279 | 77.5 | model | config |
| VGG16 | 512 | 8 | 240e | 65.975 | 80.2 | model | config |
NOTE: MobileNet-SSD is trained in 2 GPU with totoal batch size as 64 and trained 120 epoches. VGG-SSD is trained in 4 GPU with total batch size as 32 and trained 240 epoches. SSD training data augmentations: randomly color distortion, randomly cropping, randomly expansion, randomly flipping.
Face Detection
Please refer face detection models for details.
Object Detection in Open Images Dataset V5
Please refer Open Images Dataset V5 Baseline model for details.
Anchor Free Models
Please refer Anchor Free Models for details.