RetinaNet Pytorch

1. References

• [1] Focal Loss for Dense Object Detection
• [2] https://github.com/yhenon/pytorch-retinanet

2. Retina Net

2.1 Retina Net Architecture

• Backbone using all variants of DenseNet and ResNet

# investigate backbone (input shape and output shape), I will update more backbone for experiments (next step: efficient net B0-B7)
cd flame/core/model/backbone
python resnet.py --version <resnet18 -> resnet110> --pretrained <if use pretrained weight>
python densenet.py --version <densenet121 -> densenet201> --pretrained <if use pretrained weight>

• FPN

# investigate fpn (input shape and output shape)
cd flame/core/model/
python fpn.py

• Regression and Classification Head

# investigate head (input shape and output shape)
cd flame/core/model/head
python efficient_head.py
python head.py

• Anchor Generation

# investigate anchor generator (input shape and output shape)
cd flame/core/model/
python anchor_generator.py

• Retina Net: combinating all parts together including Resnet/Desenet (backbone), FPN (neck), Regressor&Classifier(head) and Anchors

2.2 Loss

• Focal Loss: for computating loss of classification head.

• Smooth L1: for computating loss of regression head.

• Focal Loss

2.3 mAP (mean Average Precision)

This is my technical note for mAP here borrowed heavily knowledge in awsome repo.

• mAP for evaluating model

2.4 Visualization

• Visualization for predicting results

3. Usage

• Training: training with Focal loss on train_set and evaluating with Focal Loss on both train_set (again) and valid_set.

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

• Evaluation: evaluating with mAP metric and visualizing all predictions for test_set.

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

4. Experiments

4.1 VOC2007 and VOC2012

<In progress ...>

4.2 COCO

<In progress ...>

4.3 Labelme format Dataset

<In progress ...>