Skip to content

Latest commit

 

History

History
140 lines (105 loc) · 4.38 KB

SOTA.md

File metadata and controls

140 lines (105 loc) · 4.38 KB
Raw

Train/Evaluate SOTA methods

  1. NetVLAD/NetVLAD+LSTM

    We develop the testbench based on Nanne/pytorch-NetVlad.

    train NetVLAD/NetVLAD+LSTM

    # NetVLAD
cd autoplace
python train.py  --nEpochs=40 --output_dim=9216 --seqLen=1 --encoder_dim=512 --num_clusters=18 --net=netvlad --split=val --logsPath=logs_netvlad --cGPU=0 --imgDir='dataset/7n5s_xy11/img'

# NetVLAD+LSTM
cd autoplace
python train.py  --nEpochs=40 --output_dim=4096 --seqLen=3 --encoder_dim=512 --num_clusters=18 --net=netvlad --split=val --logsPath=logs_netvlad --cGPU=0 --imgDir='dataset/7n5s_xy11/img'

    evaluate NetVLAD/NetVLAD+LSTM

    cd autoplace
python train.py --mode='evaluate'  --cGPU=0  --split=test --resume=[log_folder]

    calculate Recall@N and Precision-Recall: modify the path netvlad in autoplace/postprocess/parse/resume_path.json to [netvlad_log_folder], and run the command:

    cd autoplace/postprocess/parse
python parse.py --model=netvlad

  2. SeqNet

    We develop the testbench based on oravus/SeqNet. We use the pretrained NetVLAD as the black-box feature encoder to encode images to 9216D vectors, which is then downscaled to 4096D using PCA: modify the path netvlad in postprocess/parse/resume_path.json to [netvlad_log_folder] and run the command:

    cd autoplace/postprocess/parse
python parse_seqnet.py --phase='generate_features'

    use the stored 4096D vectors to train S3-SeqNet and S1-SeqNet (it can be done in parallel):

    # train S3-SeqNet
cd autoplace
python train_seqnet.py  --split=val --cGPU=0 

# train S1-SeqNet
cd autoplace
python train_seqnet.py  --split=val --seqLen=1 --w=1 --cGPU=0

    evaluate

    # evaluate S3-SeqNet or S1-SeqNet
cd autoplace
python train_seqnet.py --mode='evaluate'  --cGPU=0  --split=test --resume=[S3-SeqNet or S1-SeqNet training log folder]

    generate detailed results: modify the path s3_seqnet and s1_seqnet in postprocess/parse_logs/resume_path.json to their [log_folder] and run the command:

    cd autoplace/postprocess/parse
python parse_seqnet.py --phase='match'

  3. MinkLoc3D

    refer to minkloc3d_AutoPlace to generate minkloc3d_features.pickle

    then copy minkloc3d_features.pickle to autoplace/postprocess/parse/results

    run the command:

    cd autoplace/postprocess/parse
python parse.py --model='minkloc3d'

  4. ScanContext

    We develop the testbench based on irapkaist/scancontext.

    cd autoplace/postprocess/parse
python parse.py --model='scancontext'

  5. M2DP

    We develop the testbench based on M2DP python.

    cd autoplace/postprocess
python parse.py --model='m2dp'

  6. Under The Radar

    refer to utr_AutoPlace repo to generate utr_features.pickle.

    copy utr_features.pickle in autoplace/postprocess/parse/results

    run the command:

    cd autoplace/postprocess/parse
python parse.py --model='utr'

  7. Kidnapped Radar

    convert Cartesian images to Polar images

    cd autoplace/preprocess
python cart_to_polar.py --dataset='../dataset/7n5s_xy11'

    train

    cd autoplace
python train_kid.py  --nEpochs=80 --output_dim=32768 --seqLen=1 --encoder_dim=512 --num_clusters=64 --net=kid --logsPath=logs_kid --split=val --cGPU=0 --imgDir='dataset/7n5s_xy11/img_polar'

    evaluate

    cd autoplace
python train_kid.py --mode='evaluate'  --cGPU=0  --split=test --resume=[training log folder]

    generate detailed results

    cd autoplace/postprocess/parse
python parse.py --model='kidnapped'

Visualize Recall and Precision-Recall

After training & evaluating the above competing methods, [methods_name]_result.pickle files should be in the autoplace/postprocess/parse/results folder.

To generate (1) Reall@N curve (2) PR curve, (3) F1 Score, (4) Average Pecision, run

cd autoplace/postprocess/vis
python competing_figure.py 
python competing_scores.py