This repo implements paper Look Before You Leap: Learning Landmark Features For One-Stage Visual Grounding CVPR 2021.

Getting Started


  • python 3.7
  • pytorch 10.0
  • cuda 10.0
  • gcc 4.92 or above


  1. Then clone the repo and install dependencies.

    git clone
    cd LBYLNet
    pip install requirements.txt 
  2. You also need to install our landmark feature convolution:

    cd ext
    git clone
    cd landmarkconv/lib/layers
    python install --user
  3. We follow dataset structure DMS and FAOA. For convience, we have pack them togather, including ReferitGame, RefCOCO, RefCOCO+, RefCOCOg.

    bash data/refer/ ./data/refer
  4. download the generated index files and place them in ./data/refer. Available at [Gdrive], [One Drive] .

  5. download the pretained model of YOLOv3.

    wget -P ext

Training and Evaluation

By default, we use 2 gpus and batchsize 64 with DDP (distributed data-parallel). We have provided several configurations and training log for reproducing our results. If you want to use different hyperparameters or models, you may create configs for yourself. Here are examples:

  • For distributed training with gpus :

    CUDA_VISIBLE_DEVICES=0,1 python lbyl_lstm_referit_batch64  --workers 8 --distributed --world_size 1  --dist_url "tcp://"
  • If you use single gpu or won't use distributed training (make sure to adjust the batchsize in the corresponding config file to match your devices):

    CUDA_VISIBLE_DEVICES=0, python lbyl_lstm_referit_batch64  --workers 8
  • For evaluation:

    CUDA_VISIBLE_DEVICES=0, python lbyl_lstm_referit_batch64 --testiter 100 --split val

Trained Models

We provide the our retrained models with this re-organized codebase and provide their checkpoints and logs for reproducing the results. To use our trained models, download them from the [Gdrive] and save them into directory cache. Then the file path is expected to be <LBYLNet dir>/cache/nnet/<config>/<dataset>/<config>_100.pkl

Notice: The reproduced performances are occassionally higher or lower (within a reasonable range) than the results reported in the paper.

In this repo, we provide the peformance of our LBYL-Nets below. You can also find the details on <LBYLNet dir>/results and <LBYLNet dir>/logs.

  • Performance on ReferitGame ([email protected]).

    Dataset Langauge Split Papar Reproduce
    ReferitGame LSTM test 65.48 65.98
    BERT test 67.47 68.48
  • Performance on RefCOCO ([email protected]).

    | Dataset | Langauge | Split | Papar | Reproduce |
    | --- | --- |
    | RefCOCO | LSTM |
    | testA | 82.18 | 82.48 |
    | testB | 71.91 | 71.76 |
    | BERT |
    | testA | 82.91 | 82.82 |
    | testB | 74.15 | 72.82 |

  • Performance on RefCOCO+ ([email protected]).

    Dataset Langauge Split Papar Reproduce
    RefCOCO+ LSTM val 66.64 66.71
    testA 73.21 72.63
    testB 56.23 55.88
    BERT val 68.64 68.76
    testA 73.38 73.73
    testB 59.49 59.62
  • Performance on RefCOCOg ([email protected]).

    Dataset Langauge Split Papar Reproduce
    RefCOCOg LSTM val 58.72 60.03
    BERT val 62.70 63.20


We also provide demo scripts to test if the repo is corretly installed. After installing the repo and download the pretained weights, you should be able to use the LBYL-Net to ground your own images.


you can change the model, image or phrase in the You will see the output image in imgs/demo_out.jpg.

#!/usr/bin/env python
import cv2
import torch
from core.test.test import _visualize
from core.groundors import Net 
# pick one model
cfg_file = "lbyl_bert_unc+_batch64"
detector = Net(cfg_file, iter=100)
# inference
image = cv2.imread('imgs/demo.jpeg')
phrase = 'the green gaint'
bbox = detector(image, phrase)
_visualize(image, pred_bbox=bbox, phrase=phrase, save_path='imgs/demo_out.jpg', color=(1, 174, 245), draw_phrase=True)






This repo is organized as CornerNet-Lite and the code is partially from FAOA (e.g. data preparation) and MAttNet (e.g. LSTM). We thank for their great works.


If you use any part of this repo in your research, please cite our paper:

      title={Look Before You Leap: Learning Landmark Features for One-Stage Visual Grounding}, 
      author={Huang, Binbin and Lian, Dongze and Luo, Weixin and Gao, Shenghua},
      booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      month = {June},