We have also merged our GFR-DSOD code here into the DSOD repo at https://github.com/szq0214/DSOD. Check it out there if you interested in DSOD.

We also see some very promising results on the PASCAL VOC Comp3 Leaderboard, like https://github.com/kuangliu/torchcv. While we found they still used the ImageNet pre-trained models as the initialized parameters (kuangliu/torchcv#11). Please note that the Comp3 Challenge only allows to use the VOC12 dataset for training (without the pre-trained models). Please check your training process carefully.

If you find this helps your research, please consider citing:

@inproceedings{shenimproving,
     title={Improving Object Detection from Scratch via Gated Feature Reuse},
     author={Shen, Zhiqiang and Shi, Honghui and Yu, Jiahui and Phan, Hai and Feris, Rogerio and Cao, Liangliang and Liu, Ding and Wang, Xinchao and Huang, Thomas and Savvides, Marios}
     booktitle={The British Machine Vision Conference (BMVC)},
     year={2019}

In GFR-DSOD, we propose a recurrent feature-pyramid structure to squeeze rich spatial and semantic features into a single prediction layer that further reduces the number of parameters to learn (DSOD need learn 1/2, but GFR-DSOD need only 1/3). Thus our new model is more fit for learning from scratch, and can converge faster than DSOD. We also introduce a novel gate-controlled prediction strategy in GFR-DSOD to adaptively enhance or attenuate feature activations at different scales based on the input object size.

0. Visualizations of network structures (tools from ethereon, please ignore the warning messages):
   • [DSOD] (http://ethereon.github.io/netscope/#/gist/b17d01f3131e2a60f9057b5d3eb9e04d)
   • [GFR-DSOD] (http://ethereon.github.io/netscope/#/gist/4dcd890cb38a7f3c55a51dc1d67068a9)

Our PASCAL VOC LMDB files:

The tables below show the results on PASCAL VOC 2007, 2012 and 2012 Comp3 (training on VOC 2012 only).

PASCAL VOC test results:

Zhiqiang Shen (zhiqiangshen0214 at gmail.com)

Any comments or suggestions are welcome!