Home

Awesome

Grounding-REVERIE-Challenge

This repository is the official REVERIE Referring Expression Grounding model of REVERIE-Challenge. This code derives from UNITER, fine-tuned on REVERIE data. If you have any question, please do not hesitate to contact me, chongyang.zhao@adelaide.edu.au.

Requirements

You need Docker to run the code. For more details, please refer to UNITER for further README information. We tested on Ubuntu 20.04 and Nvidia 2080ti.

Data Preparation

Note that BBoxes_v2 files have two different forms of organization, so you need download Grounding_BBoxes_v2, which is different from the one used by HOP-REVERIE-Challenge.

Download the REVERIE data, Grounding_BBoxes_v2 and object features (reverie_obj_feats_v2.pkl or reverie_bbox_feat_v2_caffe.zip, code: nubg), both of them are extracted using BUTD Faster R-CNN trained on Visual Genome, and organise data like below:

|- Grounding-REVERIE-Challenge
    |- Downloads
        |- REVERIE
        |   |- REVERIE_val_seen.json
        |   |- REVERIE_val_unseen.json
        |   |- REVERIE_test.json
        |                     
        |- Grounding_BBoxes_v2
        |   |- 1LXtFkjw3qL_0b22fa63d0f54a529c525afbf2e8bb25.json
        |   |- 1LXtFkjw3qL_0b302846f0994ec9851862b1d317d7f2.json
        |   ...           
        |   |- zsNo4HB9uLZ_faad06c7cb2b4a6f9220e7f6f87c800b.json
        |
        |- reverie_bbox_feat_v2_caffe (option 1)
        |   |- 1LXtFkjw3qL_0b22fa63d0f54a529c525afbf2e8bb25_00.h5
        |   |- 1LXtFkjw3qL_0b22fa63d0f54a529c525afbf2e8bb25_01.h5
        |   ...           
        |   |- zsNo4HB9uLZ_faad06c7cb2b4a6f9220e7f6f87c800b_35.h5
        |
        |- reverie_obj_feats_v2.pkl (option 2)
# Image object features preparation
# Make sure the parameters in script are correct
# 
# Support .h5 and .pkl files
# .h5:
# feats_path: "Downloads/reverie_bbox_feat_v2_caffe"
# feats_format: "h5"
# .pkl:
# feats_path: "Downloads/reverie_obj_feats_v2.pkl"
# feats_format: "pkl"

bash img_data_preparation.sh

Pre-trained Model

Usage

  1.  # Docker image should be automatically pulled
    
     cd Grounding-REVERIE-Challenge
     bash launch_container.sh
    
  2.  # Object id prediction 
     # Make sure the parameters in the script are correct
     #
     # Navigation Results:
     # input_nav_dir: "input_nav_dir"
     #
     # Grounding Results (Navigation Results + predObjId):
     # output_dir: "submit_file"
     #
     # reverie_dir: "Downloads/REVERIE"
     # boxes_dir: "Downloads/Grounding_BBoxes_v2"
     #
     # We support multi-GPU inference
     # NUM_GPUS=4
    
     bash eval.sh