Testing on Standard Benchmarks

May 1, 2025 · View on GitHub

This document describes how to run MASA for the standard benchmarks testing. Our model and data can be found here: MASA.

Prepare Datasets

It is recommended to symlink the dataset root to $MASA/data, and the model root to $MASA/saved_models. If your folder structure is different, you may need to change the corresponding paths in config files.

Download TAO

a. Please follow TAO download instructions.

b. Please download converted TAO annotations and put them in data/tao/annotations/.

You can download the annotations tao_val_lvis_v05_classes.json from here.

You can download the annotations tao_val_lvis_v1_classes.json from here.

Note that the original TAO annotations has some mistakes regarding class names. We have fixed the class names in the converted annotations and make it consistent with the LVIS dataset.

Optional: generate the annotations by yourself

You can also generate the annotations by yourself. Please refer to the instructions here.

Symlink the data

Our folder structure follows

├── masa
├── tools
├── configs
├── results
├── data
    ├── tao
        ├── frames
            ├── train
            ├── val
            ├── test
        ├── annotations

|── saved_models # saved_models are the folder to save downloaded pretrained models and also the models you trained.
    ├── pretrain_weights
    ├── masa_models

It will be easier if you create the same folder structure.

Download BDD100K

We present an example based on BDD100K dataset. Please first download the images and annotations from the official website.

On the download page, the required data and annotations are

mot set images: MOT 2020 Images
mot set annotations: MOT 2020 Labels
mots set images: MOTS 2020 Images
mots set annotations: MOTS 2020 Labels

Symlink the data

It is recommended to symlink the dataset root to $MASA/data. The official BDD100K annotations are in the format of scalabel. Please put the scalabel annotations file udner the scalabel_gt folder. Our folder structure follows

├── masa
├── tools
├── configs
├── results
├── data
│   ├── bdd
│   │   ├── bdd100k
            ├── images  
                ├── track 
                    |── val
        ├── annotations 
        │   ├── box_track_20
        │   ├── det_20
        │   ├── scalabel_gt
                |── box_track_20
                    |── val
                |── seg_track_20
                    |── val

Convert annotations to COCO style

The official BDD100K annotations are in the format of scalabel.

You can directly download the converted annotations: mot and mots and put them in the data/bdd/annotations/ folder.

(Optional) If you want to convert the annotations by yourself, you can use bdd100k toolkit. Please install the bdd100k toolkit by following the instructions here. Then, you can run the following commands to convert the annotations to COCO format.

mkdir data/bdd/annotations/box_track_20
python -m bdd100k.label.to_coco -m box_track -i data/bdd/annotations/scalabel_gt/box_track_20/${SET_NAME} -o data/bdd/annotations/box_track_20/bdd_box_track_${SET_NAME}_cocofmt.json

The ${SET_NAME} here can be one of ['train', 'val', 'test'].

Download the public detections

Create a folder named 'results' under the root.

mkdir results

Download the public detections from here and unzip it under the 'results' folder.

Run MASA

This codebase is inherited from mmdetection. You can refer to the offical instructions. You can also refer to the short instructions below. We provide config files in configs.

Test a model with COCO-format

Note that, in this repo, the evaluation metrics are computed with COCO-format. But to report the results on BDD100K, evaluating with BDD100K-format is required.

single GPU
single node multiple GPU

You can use the following commands to test a dataset.

# single-gpu testing
python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--cfg-options]

# multi-gpu testing
tools/dist_test.sh ${CONFIG_FILE} ${CHECKPOINT_FILE} ${GPU_NUM} [--cfg-options]

Optional arguments:

--cfg-options: If specified, some setting in the used config will be overridden.

Test on TAO TETA benchmark

We provide the config file for testing on the TAO TETA benchmark. Use MASA-GroundingDINO for example.

tools/dist_test.sh configs/masa-gdino/tao_teta_test/masa_gdino_swinb_tao_test_detic_dets.py saved_models/masa_models/gdino_masa.pth 8

Test on Open-vocabulary MOT benchmark

We provide the config file for testing on the open-vocabulary MOT benchmark. Use MASA-Detic for example.

tools/dist_test.sh configs/masa-detic/open_vocabulary_mot_test/masa_detic_swinb_open_vocabulary_test.py saved_models/masa_models/detic_masa.pth 8

Test on BDD100K MOT benchmark

We provide the config file for testing on the BDD100K MOT benchmark, with ByteTrack's YoloX prediction. Use MASA-GroundingDINO for example.

tools/dist_test.sh configs/masa-gdino/bdd_test/masa_gdino_bdd_mot_test.py saved_models/masa_models/gdino_masa.pth 8

Test on BDD100K MOTS benchmark

We provide the config file for testing on the BDD100K MOTS benchmark, with UNINEXT's prediction. Use MASA-GroundingDINO for example.

tools/dist_test.sh configs/masa-gdino/bdd_test/masa_gdino_bdd_mots_test.py saved_models/masa_models/gdino_masa.pth 8

For other models, you can replace the ${CONFIG_FILE} and ${CHECKPOINT_FILE} with the corresponding paths.