Image Retrieval

北京交通大学2024春季《计算机视觉基础》

交大视觉印象数据集2024

1. Prepare training dataset

Put the data under ${DATA_ROOT}. The prepared directory would look like:

${DATA_ROOT}
├── base
│   ├── fh
│   ├── mh
│   ...
│   ├── zx
│   ├── util_pic
├── query
│   ├── fh
│   ├── mh
│   ...
│   ├── zx

${DATA_ROOT} is set to ./data by default, which can be modified via hydra command line interface --data '/your/data/path'.

Run split.py first to split the images in util_pic by label (totally 17 classes). Then delete util_pic.

Now we have a training dataset (is set to ./data/base by default) containing 6445 images of 25 categories.

2. Fine-tuning

• Prepare the environment

conda activate your_enviroment
cd /your/path/to/your_project

• AlexNet w/o latent layer

python finetune.py --model alexnet --batchsize 64 --lr 0.001 --num_epochs 300 --data './your/data/path' --seed 42

• AlexNet w/ latent layer CVPRWorkshop2015 Deep Learning of Binary Hash Codes for Fast Image Retrieval

python finetune.py --model alexnet --batchsize 64 --lr 0.001 --num_epochs 300 --data './your/data/path' --seed 42 --latent_layer --latent_size 48

• ResNet-50 w/o latent layer

python finetune.py --model resnet --batchsize 64 --lr 0.001 --num_epochs 300 --data './your/data/path' --seed 42

• Memory Usage

  Running on NVIDIA GeForce RTX 3090.

  

3. Retrieval

Adding --plot to the command line will export the retrieved images for each image in the query set. You can find them in ./plots/your_model_name/[20/40/60].

• AlexNet w/o latent layer

python retrieval.py --model alexnet --data './your/data/path' --dist [cos/euclidean] [--plot]

• AlexNet w/ latent layer

python retrieval.py --model alexnet --data './your/data/path' --latent_layer --dist [cos/euclidean] [--plot]

• ResNet-50 w/o latent layer

python retrieval.py --model resnet --data './your/data/path' --dist [cos/euclidean] [--plot]

Example of results

• AlexNet w/o latent layer (cosine_similarity)

  08.png in ./plots/your_model_name/20.

  

  [email protected] in ./plots/your_model_name.

  

• AlexNet w/o latent layer (euclidean_dist)

  

• AlexNet w/ latent layer (binary + cosine_similarity)

  08.png in ./plots/your_model_name/20.

  

  [email protected] in ./plots/your_model_name.

  

• AlexNet w/ latent layer (binary + euclidean_dist)

  

• ResNet-50 w/o latent layer (cosine_similarity)

  16.png in ./plots/your_model_name/60.

  

  [email protected] in ./plots/your_model_name.

  

• ResNet-50 w/o latent layer (euclidean_dist)