Efficient Det: Scalable and Efficient Object Detection

[1] EfficientDet - Scalable and Efficient Object Detection: https://arxiv.org/pdf/1911.09070.pdf
[2] EfficientDet - zylo117 Github: https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch
[3] EfficinetNet - Github: https://github.com/lukemelas/EfficientNet-PyTorch

TODO

• Dataset format including coco, pascal, labelme and altheia.
• Efficient Det training, inference flow.
• Add mask head, based on the idea of RetinaMask
• Augmentations: Mosaic, Mixup, CutMix,...
• BBox IOU Loss (GIoU, DIoU, CIoU, ...)

• Distance-IoU Loss
• Complete-IoU Loss

• Using soft-nms instead of normal NMS.
• FP16 (automatic mixed precision)
• DDP (DistributedDataParallel) for faster training on GPUs.
• Tensorboard, Logger.
• Profiler.

MAIN FUNCTIONS

• Backbone using all variants of Efficient NetV1
• BiFPN
• Regression and Classification Head
• Anchor Generation
• Efficient Det: combinating all parts together including Efficient NetV1 (backbone), BiFPN (neck), Regressor&Classifier(head) and Anchors
• Focal Loss
• COCO Eval
• mAP
• Visualization for predicting results

DATASET

ID	Dataset Name	Train	Val	Test	Format
1	COCO 2017	118,287	5,000	-	COCO JSON
2	Pascal VOC 2007	5,011	4,952	-	PASCAL XML
3	Pascal VOC 2012	1,464	1,449	-	PASCAL XML
4	PubLayNet	335,703	11,245	11,405	COCO JSON

• COCO Train/Val/Test 2017

https://cocodataset.org/#download

• Pascal VOC 2007

http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar

http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

• Pascal VOC 2012

http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar

• Publaynet

https://developer.ibm.com/exchanges/data/all/publaynet/

Note. Run TensorBoard on Server

You have to create a ssh connection using port forwarding:

ssh -L 16006:127.0.0.1:6006 user@host

Then you run the tensorboard command:

tensorboard --logdir=/path/to/logs

Then you can easily access the tensorboard in your browser under:

localhost:16006/

USAGE

VOC 2007, 2012

• Config: https://github.com/phungpx/efficient_det_pytorch/tree/master/configs/PASCAL
• Training

CUDA_VISIBLE_DEVICES=<cuda_indice> python -m flame configs/PASCAL/pascal_training.yaml

• Testing

CUDA_VISIBLE_DEVICES=<cuda_indice> python -m flame configs/PASCAL/pascal_testing.yaml

• Result

Model	Parameters	Result
EffiDet - D0	3,839,117
EffiDet - D1	-	-
EffiDet - D2	-	-
EffiDet - D3	-	-
EffiDet - D4	-	-
EffiDet - D5	-	-
EffiDet - D6	-	-
EffiDet - D7	-	-
EffiDet - D7x	-	-

COCO

• Config: https://github.com/phungpx/efficient_det_pytorch/tree/master/configs/COCO/
• Training

CUDA_VISIBLE_DEVICES=<cuda_indice> python -m flame configs/COCO/coco_training.yaml

• Testing

CUDA_VISIBLE_DEVICES=<cuda_indice> python -m flame configs/COCO/coco_testing.yaml

• Result

Model	Parameters	Result
EffiDet - D0	3,874,217
EffiDet - D1	-	-
EffiDet - D2	-	-
EffiDet - D3	-	-
EffiDet - D4	-	-
EffiDet - D5	-	-
EffiDet - D6	-	-
EffiDet - D7	-	-
EffiDet - D7x	-	-

Birdviews