DSSD MXNET

Features

• An MXNet implementation of DSSD : Deconvolutional Single Shot Detector.

• MXNET implementation for TDM architecture,Beyond Skip Connections Top Down Modulation for Object Detection.

• Code implementation in this repository is based on the MXNET implementation of SSD available here.

Specific VOC mAP Results

| backbone | SSD | SSD_min_loss | DSSD* | DSSD_stage1 | DSSD_stage2 | SSD+TDM | | ------------- | ----- | ------------ | --------- | ----------- | ----------- | --------- | | vgg16-512 | 75.56 | 75.35 | 75.85 | 65.16 | 75.77 | 76.80 | | vgg16-300 | 74.65 | 74.59 | 75.74 | * | * | * | | resnet101-512 | 78.43 | 78.02 | *79.25 | 71.98 | 78.19 | 79.18 | | resnet101-321 | 75.18 | 74.80 | 75.54 | * | * | ** |

• DSSD*: results by use our traning strategy,for more details,please see here

Requirements

We tested our code on:

Ubuntu 16.04, Python 2.7 with

numpy(1.11.0), cv2(3.3.0-dev)

mxnet 0.11.0

Preparation for Training

1.Download the converted pretrained vgg16_reduced model here.

2.Prepare VOC datasets and generate .rec files by using tools/prepare_pascal.sh

3.Set TDM or DSSD mode in function get_config from symbol/symbol_factory.py.By default, DSSD mode is used,please set all configs by your needs.

4.start training

python train.py

or choice a bash file which provide in ./script/ to run some default parameters setting,such like try the two stage training strategy.

bash scripts/stage1_dssd_train_res_voc.sh

Demo

1. Download model, available at here, and place it in the model folder.

2. run demo.py

References

1.SSD: Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu CY, Berg AC. SSD: Single shot multibox detector. InEuropean Conference on Computer Vision 2016 Oct 8 (pp. 21-37). Springer International Publishing.Link

2.DSSD: Fu CY, Liu W, Ranga A, Tyagi A, Berg AC. DSSD: Deconvolutional Single Shot Detector. arXiv preprint arXiv:1701.06659. 2017 Jan 23. Link

3.TDM: Shrivastava A, Sukthankar R, Malik J, Gupta A. Beyond Skip Connections: Top-Down Modulation for ObjectDetection. arXiv preprint arXiv:1612.06851. 2016 Dec 20.Link