Skip to content

Latest commit

 

History

History
72 lines (61 loc) · 3.28 KB

GETTING_STARTED.md

File metadata and controls

72 lines (61 loc) · 3.28 KB
Raw

Getting started

LOD requires running multiple steps. This instruction will guide you through these steps and run LOD on C20K dataset with VGG16 features. Please make sure that you prepare the necessary data by following DATA.md.

We use SLURM to launch distributed jobs. The launching slurm sbatch files and necessary scripts can be found in slurm folder. To customize to another datasets or using a different number of CPUs/jobs, you can edit the parameters in these files. You might need to customize SLURM parameters (e.g., memory/time limit) in .s files to run on your own datasets.

For all of the following steps, you are supposed to be in the slurm/vgg16 folder.

1. Image feature extraction

The first step of LOD involves extracting image features for all images in the dataset. The features will then be used to generate region proposals and compute image initial neighbors. In Bash, run:

# The default number of CPUs/jobs is 200
# Expected time: 450 seconds / job
bash run_extract_image_feat.sh

After all jobs are finished, you will have NUM_IMAGES files numbered from 1.mat to <NUM_IMAGES>.mat in subfolders vgg16_relu43, vgg16_relu53 and vgg16_fc6_resize of data/<DATASET_NAME>/mixed/features/image.

2. Proposal generation

We use the method from rOSD to generate proposals from CNN features. In Bash, run:

# The default number of CPUs/jobs is 200
# Expected time: 450 seconds / job
bash run_create_proposals.sh

After all jobs are finished, we gather their outputs to obtain the proposal set coco_train_20k_vgg16_cnn_03_20_10_05_10_with_roots_2000. In Matlab, run:

run_proposals_processing;

3. Initial image neighbor computation

In Matlab run:

run_initial_image_neighbor_computation;

This script will print the number of score blocks (S_ij) to compute. You should note this down. It will be a parameter to launch jobs in the Score computation step.

4. Region feature extraction

In Bash, run:

# The default number of CPUs/jobs is 200
# Expected time: 420 seconds / job
bash run_extract_roi_feat.sh

After all jobs are finished, you will have NUM_IMAGES files numbered from 1.mat to <NUM_IMAGES>.mat in subfolder vgg16_relu53_77_roi_pooling_noresize of data/<PROPSET_NAME>/mixed/features/proposals.

5. Score computation

LOD uses PHM score. To compute scores in parallel with slurm, in Bash, run:

# The default number of CPUs/jobs is 1200
# Expected time: 1 hour 30 minutes / job
bash run_compute_scores.sh <Number of blocks at step 3 here>

After all jobs are finished, in Matlab, run the following code to generate chunks of the score matrix:

run_scores_processing;

6. Ranking

In Matlab, run:

# Expected time: 45 minutes
run_ranking;

After the optimization is finished, the proposals' importance scores are saved in data/coco_train_20k_vgg16_cnn_03_20_10_05_10_with_roots_2000/mixed/lsuod_results/em_ppr_SSMC

7. Obtain objects and evaluation

In Matlab, run:

run_ranking_processing_and_evaluation;