Awesome
mxnet-dsod
Under development.This is re-implementations of DSOD and grp-DSOD,which train object detector from scratch. More informations from DSOD: Learning Deeply Supervised Object Detectors from Scratch and Learning Object Detectors from Scratch with Gated Recurrent Feature Pyramids
Prerequisites
- Python 3.6
- Mxnet
- Numpy
- Opencv-python
Preparations
- Clone this repository.
- Download VOC dataset from this released page.Make .rec file using im2rec.py and put them in folder data if you want to train on VOC.A tutorial
Train the model
# train dsod
python train.py
# train grp-dsod
python train.py --network grpdsod --data-shape 320
# see advanced arguments for training
python train.py -h
TODO
- Mutil-GPUs support.I just have one gpu now,so I don't konw whether the Synchronized Batch Normalization works well or not.I will updata it when I can use more gpu.Or you can modify it by youself.
- A pretrained model.DSOD and Grp-dsod converge very slowly,the author gets a good performance by training them 100000 epochs on VOC.It will take a long time.And I will train it if my gpu is free.Don't expect too much.
- Training on a small dataset.