Skip to content

Latest commit

 

History

History
135 lines (95 loc) · 2.89 KB

README_paper.md

File metadata and controls

135 lines (95 loc) · 2.89 KB
Raw

How to reproduce the paper results.

wasabi

Image-Based Place Recognition on Bucolic Environment Across Seasons From Semantic Edge Description

Experiments code (coming soon, soon being in the order of days)

Setup

git clone --recursive 
pip3 install -r requirements.txt

Datasets

CMU-Seasons

Script to download slice [6,8] and [22,25]:

./get_cmu.sh

Symphony

Script to Symphony images:

./get_symphony.sh

WASABI Results

./scripts/wasabi.sh <trial>
python -m plots.tab2latex --method wasabi --trial <trial> --data cmu_park
pdflatex res/wasabi/1/1.tex

NetVLAD Results

./scripts/netvlad.sh <trial>
python -m plots.tab2latex --method netvlad --trial <trial> --data cmu_park

VLAD/BOW Results

With original codebook trained on Flickr100

./scripts/vlad_bow.sh <trial> <agg_mode> flickr100k

  # example (See scripts/vlad_bow.sh to run it)
  python3 -m methods.vlad.retrieve \
    --trial 0 \
    --dist_pos 5 \
    --top_k 20 \
    --lf_mode sift \
    --max_num_feat 1000 \
    --agg_mode bow \
    --centroids flickr100k \
    --vlad_norm ssr \
    --data cmu \
    --slice_id 22 \
    --cam_id 0 \
    --survey_id 0 \
    --img_dir "$img_dir" \
    --meta_dir "$meta_dir" \
    --n_words 64

Generate codebook on CMU-Seasons/Lake

With codebook finetuned on cmu park

With codebook finetuned on lake

DELF results

Install DELF

git submodule update --init third_party/models

Here is a variation of the delf installation guidelines. The only variations concerns the paths.

Protobuf

Follow [https://github.com/tensorflow/models/blob/master/research/delf/INSTALL_INSTRUCTIONS.md#protobuf](this link).

tensorflow/models

# First, install slim's "nets" package.
cd third_party/models/research/slim/
pip3 install -e .

Then, compile DELF's protobufs. Use PATH_TO_PROTOC as the directory where you downloaded the protoc compiler.

# From third_party/models/research/delf/
${PATH_TO_PROTOC?}/bin/protoc delf/protos/*.proto --python_out=.

Finally, install the DELF package.

# From third_party/models/research/delf/
pip3 install -e . # Install "delf" package.

Update your python paths:

export PYTHONPATH=$PYTHONPATH:<full path to>/third_party/models/research:<full path to>/third_party/models/research/slim

At this point, running

python3 -c 'import delf'

should just return without complaints. This indicates that the DELF package is loaded successfully.

Convert DELF codebooks from tf to np format

    cd meta/words/delf
    ./get_words.sh
    ./convert_words.sh

    cd "$WASABI_DIR"
    python3 methods/delf/codebook_tf2np.py \
        --ckpt_dir meta/words/delf/roxford5k_codebook_65536/k65536_codebook_tfckpt/ \
        --out meta/words/delf/oxford5k_65536.txt

Retrieval

./scripts/delf.sh

# Paper plots