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

1. Python 3.6
2. Mxnet
3. Numpy
4. Opencv-python

Preparations

1. Clone this repository.
2. 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

1. 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.
2. 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.
3. Training on a small dataset.