/ZSD_Release

This repository provides evaluation code for paper titled "Zero-Shot Object Detection: Learning to Simultaneously Recognize and Localize Novel Concepts."

Primary LanguagePythonMIT LicenseMIT

Zero-Shot Object Detection

This code is the testing implementation of the following work:

Shafin Rahman, Salman Khan, and Fatih Porikli. "Zero-Shot Object Detection: Learning to Simultaneously Recognize and Localize Novel Concepts." arXiv preprint arXiv:1803.06049 (2018). (Project Page)

ProblemFigure

Requirements

  • Download the pre-trained model available on the link below and place it inside the "Model" directory (Link to pre-trained model (h5 format))
  • Other requirements:
    • Python 2.7
    • Keras 2.1.4
    • OpenCV 2.4.13
    • Tensorflow 1.3.0

This code has also been tested with Python 3.6, Keras 2.0.8, OpenCV 3.4.0 and on Ubuntu and Windows.

Files

sample_input.txt: a sample input file containing test image paths detect.py: to perform zero-shot detection task using sample_input.txt keras_frcnn: directory containing supporting code of the model Dataset: directory containing sample input and output Model: directory containing pre-trained model and configuration file ImageNet2017

  • cls_names.txt: list of 200 class names of ImageNet detection dataset.
    • seen.txt: list of seen class names used in the paper
    • unseen.txt: list of unseen class names used in the paper
    • train_seen_all.zip: it is a zipped version of text file train_seen_all.txt which contain training image paths and annotation used in the paper. Each line contains training image path, a bounding box co-ordinate and the ground truth class name of that bounding box. For example, Filepath,x1,y1,x2,y2,class_name
    • unseen_test.txt: all the image paths used for testing in the papers. Images are from training and validation set from ImageNet 2017 detection challenge. Every image contains at least one instance of unseen object.
    • word_w2v.txt: word2vec word vectors of 200 classes + bg used in the paper. The ith column represents the 500-dimensional word vectors of the class name of ith row of cls_names.txt.
    • word_glo.txt: GloVe word vectors of 200 classes + background (bg) used in the paper. The ith column represents the 300-dimensional word vectors of the class name of ith row of cls_names.txt.

Running instruction

To run zero-shot detection on sample input kept in Dataset/Sampleinput, simply run detect.py after installing all dependencies like Keras, Tensorflow, OpenCV and placing the pre-trained model in the Model directory. This code will generate the output files for each input image to Dataset/Sampleoutput.

Notes on ImageNet experiments

The resources required to reproduce results of ImageNet related experiments are kept in the directory ImageNet2017. All the images are from ILSVRC2017_DET.tar.gz which can be obtained from ImageNet detection challenge 2017 website. For both training and testing of this paper, we have used images from /ILSVRC/Data/DET/train and /ILSVRC/Data/DET/val of the zipped arxiv ILSVRC2017_DET.tar.gz.

Trubleshooting

  • If you get the CUDA_ERROR_OUT_OF_MEMORY in Tensorflow, place the following snippet after library loadings (the top section) in detect.py

     import tensorflow as tf 
 config = tf.ConfigProto()
 config.gpu_options.allow_growth = True
 config.gpu_options.per_process_gpu_memory_fraction = 0.8 #Change it to suit your GPU load
 K.set_session(tf.Session(config=config))

  • If you want to run the sample code on CPU only, place the following snippet after library loadings (the top section) in detect.py

     import tensorflow as tf 
 num_cores = 2 # 2,4, or 8
 config = tf.ConfigProto(intra_op_parallelism_threads=num_cores,\
     inter_op_parallelism_threads=num_cores, allow_soft_placement=True,\
     device_count = {'CPU' : 1, 'GPU' : 0})
 K.set_session(tf.Session(config=config))

Citation

Citation

If you use this code and model for your research, please consider citing:

@article{rahman2018zeroshot, 
title={Zero-Shot Object Detection: Learning to Simultaneously Recognize and Localize Novel Concepts}, 
author={Rahman, Shafin and Khan, Salman and Porikli, Fatih}, 
journal={arXiv preprint arXiv:1803.06049}, 
year={2018} 
}

Acknowledgment

We thank Yann Henon for the following implementation of Faster-RCNN: keras-frcnn