This is a Matlab implementation for our paper:
Spatial-Content Image Search in Complex Scenes.
Note: This repository may not contain every single step of our algorithm, some codes are lost due to my graduation from university. However, core function codes are still here. Sorry for that.
- Library yael: Most functions are already contained in folder
/utils
. Other functions needed could be found in this library.
- Detect boxes using YOLOv3 (not include).
- Extract image features using Googlenet. In this paper, we use its Caffe implementation (not include).
- Build image representation using
preprocess_data.m
. - Image search with our spatial-content similarity
test_spacon.m
.
extract_text.m
extract annotations from coco-api, then use./pycode/create_tfidf.py
to calculate standard relevance score. As shown in paper 3.1.modify_bboxes.m
change the format of original bounding boxes detected by yolov3.postprocess.m
L2 normalization.spaconsim.m
spatial-content similarity of two image.trans_ind.m
a small function used in calculating one object's visual feature.