cedrickchee / ssd-yolo-retinanet

Multi-class object detection pipeline—Single Shot MultiBox Detector (SSD) + YOLOv3 (real-time) + focal loss (RetinaNet) + Pascal VOC 2007 dataset

Realtime Multi-object Detection Pipeline

Note: this repo is currently under heavy development. It's not ready for general consumption. So, please refrain yourself from using it in production.

The goal of this project is to buid a single end-to-end deep learning model for more accurate and faster (near real-time) multi-object detection that can be train in single-pass of multiple different pieces:

• Single Shot MultiBox Detector (SSD)
• YOLOv3 real-time properties
• Focal loss for dense object detection (RetinaNet)
• Non Maximum Suppression (NMS)
• Scalable object detection using deep neural networks
• Faster R-CNN tricks

These techniques and methods from various research papers will be implemented using PyTorch.

We will be using Pascal VOC2007 dataset.

Requirements

• Python 3
• Pytorch 0.4
• numpy
• fastai PyTorch library

Training

# Select the script that you want to train for reproducing a results
./retina_ce_sgd_0.001.sh
# For the focal loss use ./retina_focal_sgd_0.0001.sh

You can see the details in trainer.py

VOC Dataset

Download VOC2007 trainval & test

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>

Download VOC2012 trainval

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>

TODO

• Build SSD + YOLO model
• Apply cross entropy loss and focal loss
• Compare between CE loss and focal loss
• Report results on VOC
  • currently achieved 50mAP on VOC2007.
• Report results on COCO
• Use relative path for easy reproducing of result

ssd-yolo-retinanet's Issues

Hi

Does the "50%mAP on VOC2007" you mentioned in TODO refer to YOLOV3?

OHEM and focal loss, which is better?

Hi, is the focal loss works in SSD? How much improvement is reached in VOC2007 evaluation when SSD is trained with focal loss? I want to know whether the focal loss performs better than OHEM(used in original SSD). Thank you!