In current referring expression tasks, most of work take the maskrcnn pretrained on MSCOCO-refervaltest dataset as object detector to extract a set of object proposals. In this setting, the object detector can provide nearly ground-truth proposal during training stage.
In additional, the repo maskrcnn provide by lichengunc is out-of-data, which was written by python2 and low PyTorch version. To facilitate the development of the referring expression reasoning community, we provide the new pretrained object detector model based on the Detectron2. I hope this will help all researchers to develop faster, accurate referring reasoning system.
| Backbone | training set | excluded images | Box-AP/AP50/AP75 | mask-AP/AP50/AP75 | Model | Log |
|---|---|---|---|---|---|---|
| ResNet101-C4 | COCO2017train | refcoco&refcoco+ val&test | 41.01/60.34/44.18 | 35.33/56.98/37.49 | ckpt_final | log |
| ResNet101-C4 | COCO2017train | refcocog val&test | 40.99/60.15/44.15 | 35.33/57.03/37.75 | ckpt_final | log |
| ResNet101-C4 | COCO2017train | refcoco&+&g val&test | 42.08/61.60/45.45 | 36.36/58.35/38.89 | ckpt_final | log |
MSCOCO-refervaltest denotes MSCOCO2017 training images minus the refcoco, refcoco+, refcocog val&test images
- Detectron2 Compiler GCC 7.4
- PyTorch 1.4.0+cu100
- Pillow 7.1.2
- cv2 4.2.0
See INSTALL.md.