Introduction

In this repo, Fast Moving Object (FMO) datasets (FMOv2, TbD-3D, TbD and Falling Objects, all available at https://cmp.felk.cvut.cz/fmo/) with additional ground truth information in JSON format (our new metadata is called FMOX) are provided and used for benchmarking trackers.

📌 If you are using this repo in your research or applications, please cite our papers related to this work (published in IMVIP 2025 and AICS 2025).

Benchmarking SAM2-based trackers on FMOX (AICS 2025)

[Paper AICS 2025] [Code] [Arxiv]

Several object tracking pipelines extending Segment Anything Model 2 (SAM2) have been proposed in the past year, where the approach is to follow and segment the object from a single exemplar template provided by the user on a initialization frame. We propose to benchmark these high performing trackers (SAM2, EfficientTAM, DAM4SAM and SAMURAI) on datasets containing fast moving objects (FMO) specifically designed to be challenging for tracking approaches. The goal is to understand better current limitations in state-of-the-art trackers by providing more detailed insights on the behavior of these trackers. We show that overall the trackers DAM4SAM and SAMURAI perform well on more challenging sequences.

@inproceedings{AICS2025-Aktas,
title={Benchmarking SAM2-based Trackers on FMOX},
author={Senem Aktas and Charles Markham and John McDonald and Rozenn Dahyot},
booktitle={33rd International Conference on Artificial Intelligence and Cognitive Science (AICS 2025)},
address={Dublin, Ireland},
year={2025},
month={December},
pages={1-12},
doi={10.48550/arXiv.2512.09633},
url={Paper=https://cvmlmu.github.io/FMOX/FMOXaics2025.pdf ArXiv=https://arxiv.org/abs/2512.09633 
Code=https://cvmlmu.github.io/FMOX/},
abstract={},
keywords={},
note={},
}

R code AICS2025.qmd (that creates AICS2025.html) used for drawing some figures in the paper is also provided in this repo.

Benchmarking EfficientTAM on FMO datasets (IMVIP 2025)

[Paper IMVIP 2025] [Code] [Arxiv]

In this repo, we extend Fast Moving Object (FMO) datasets (FMOv2, TbD-3D, TbD and Falling Objects, all available at https://cmp.felk.cvut.cz/fmo/) with additional ground truth information in JSON format (our new metadata is called FMOX). The provided FMOX JSON format allows for seamless compatibility with various machine learning frameworks, making it easier for developers and researchers to utilize the dataset in their applications. With FMOX, we test a recently proposed foundational model for tracking (EfficientTAM) showing that its performance compares well with the pipelines originally developed for these FMO datasets.

Scripts provided in this repo allow to download all FMO datasets, create json metadata, and assess object tracking with EfficientTAM using with TIoU metric.

@inproceedings{FMOX_AKTAS2025,
  title={Benchmarking EfficientTAM on FMO datasets},
  author={Senem Aktas and Charles Markham and John McDonald and Rozenn Dahyot},
  booktitle={Irish Machine Vision and Image Processing},
  doi={10.48550/arXiv.2509.06536},
  url={https://cvmlmu.github.io/FMOX/},
  month={September},
  pages={59-66},
  address={Ulster University, Derry-Londonderry, Northern Ireland},
  year={2025},
}

Installation

Getting started

git clone https://github.com/CVMLmu/FMOX.git branch main
cd FMOX
# for conda, create the environment using:
conda env create -n fmo_data_env -f environment.yml
conda activate fmo_data_env

Notebooks

The following notebooks can be run in that environment:

• create_jsons_main.ipynb or main.py for creating FMOX JSON files in folder FMOX-Jsons (these are already provided in this Github).
• fmox_main.ipynb or fmox_main.py to evaluate efficientTAM on FMOX with TIOU (i.e. all 4 FMO datasets using our JSON metadata files). EfficientTAM results are already computed and are provided in JSON in the folder EfficientTAM-Jsons: we have used the pretrained model EfficientTAM-S with its default parameters.
• Note: Downloading the dataset (except MATLAB files) may take some time, as all downloads are approximately 24 GB in size. If the dataset already exists in the specified location, the code will automatically skip this step.

Repo tree structure

  environment.yml
│   LICENSE
│   README.md
│
└───FMOX-code
    │   download_datasets.py
    │   __init__.py
    │
    ├───create-FMOX
    │   │   combine_all_mask_to_single_img.py
    │   │   create_fmov2_json.py
    │   │   create_jsons_main.ipynb
    │   │   create_tbd_json.py
    │   │   main.py
    │   │   rle_to_seg_mask_img.py
    │   │   tbd_visualize_bboxes.py
    │   │
    │   └───dataset_loader
    │           create_json_via_benchmark_loader.py
    │           loaders_helpers.py
    │           reporters.py
    │
    ├───EfficientTAM-Jsons
    │       efficienttam_All4.json
    │       efficienttam_falling.json
    │       efficientTam_fmov2.json
    │       efficienttam_tbd3d.json
    │       efficienttam_tdb.json
    │
    ├───FMOX-Jsons
    │       FMOX_All4.json
    │       FMOX_fall_and_tbd3d.json
    │       FMOX_fmov2.json
    │       FMOX_tbd.json
    │       FMOX_tbd_whole_sequence.json
    │
    └───use-FMOX
        │   access_json_bboxes.py
        │   calciou.py
        │   csv_to_graphics.py
        │   efficientam_evaluation.py
        │   EfficientTAM_averageTIoU.csv
        │   FMOX_all4_json_to_CSV.py
        │   FMOX_All4_statistics.csv
        │   fmox_main.ipynb
        │   fmox_main.py
        │   size_label_bar.png
        │   size_label_count.py
        │   vis_trajectory.py
        │   __init__.py
        │
        └───efficientTAM_traj_vis
                efficientTAM_traj_Falling_Object_v_box_GTgamma.jpg
                (...)

• FMOX-Jsons folder contains all metadata files contains for the FMO datasets (created with create_jsons_main.ipynb):

  • FMOX_All4.json: All Falling Object, FMOV2, TbD and TbD-3D dataset annotations are in that JSON file.
  • FMOX_fall_and_tbd3d.json: Falling Object and TbD-3D dataset annotations are in that JSON file.
  • FMOX_fmov2.json : Only FMOV2 dataset annotations in a JSON file due to annotation format differences
  • FMOX_tbd.json: TbD annotations in a JSON file, annotations obtained with fall_and_tbd3d datasets.
  • FMOX_tbd_whole_sequence.json : Only TbD dataset annotations in a JSON file due to annotation format differences.

  The JSON files include the ground truth bounding boxes for the four datasets examined. Note that there are two ways to create the TbD JSON. The first method involves creating FMOX_tbd.json, which is obtained using the same code (create-FMOX/dataset_loader) as fall_and_tbd3d.json. The second method involves creating FMOX_tbd_whole_sequence.json, which contains bounding boxes for the entire sequence, not just for the FMOs and obtained directly from the provided original ground truth text files.

• EfficientTAM-Jsons: This folder contains the evaluation results of EfficientTAM saved in FMOX JSON format. After installing the EfficientTAM repository, the first bounding box coordinates from the FMOX JSON files for each sequence are provided to initialize the EfficientTAM tracker.

Additional Information

The following results are shared in this repo (created with fmox_main.ipynb):

• FMOX_All4_statistics.csv: This describes all the sequences from the 4 FMO datasets.
• EfficientTAM_averageTIoU.csv: This provides the TIoU obtained with EfficientTAM.

FMOX Object Size Categories

The sizes of the objects in the public FMO datasets were calculated and "object size levels" were assigned. A total of five distinct level defined as below:

Extremely Tiny	Tiny	Small	Medium	Large
[1 × 1, 8 × 8)	[8 × 8, 16 × 16)	[16 × 16, 32 × 32)	[32 × 32, 96 × 96)	[96 × 96, ∞)

Table: FMOX object size categories.

Structure of FMOX

{
  "databases": [
    {
      "dataset_name": "Falling_Object",
      "version": "1.0",
      "description": "Falling_Object annotations.",
      "sub_datasets": [
        {
          "subdb_name": "v_box_GTgamma",
          "images": [
            {
              "img_index": 1,
              "image_file_name": "00000027.png",
              "annotations": [
                {
                  "bbox_xyxy": [161, 259, 245, 333],
                  "object_wh": [84, 74],
                  "size_category": "medium"
                }
              ]
            },
            {
              "img_index": 2,
              "image_file_name": "00000028.png",
              "annotations": ["bbox_xyxy": [.....], "object_wh": [.....], "size_category": "...." ]
            }
          ]
        }
      ]
    }
  ]
}

Average TIoU Performance Comparison

This table compares the average TIoU $(\uparrow)$ performance of various studies on FMO datasets. Evaluation of the EfficientTAM has been done with FMOX Json. The best results are indicated with $^*$ and the second-best results with $^{**}$.

Datasets	Defmo [Rozumnyi et al., 2021]	FmoDetect [Rozumnyi et al., 2021]	TbD [Kotera et al., 2019]	TbD-3D [Rozumnyi et al., 2020]	EfficientTAM [Xiong et al., 2024]
Falling Object	0.684**	N/A	0.539	0.539	0.7093*
TbD	0.550**	(a) 0.519 (b) 0.715*	0.542	0.542	0.4546
TbD-3D	0.879*	N/A	0.598	0.598	0.8604**

(a) Real-time with trajectories estimated by the network. (b) With the proposed deblurring. N/A indicates "Not defined".

Acknowledgments

This Github repo was created and developped by Senem Aktas, and in addition it was tested by Rozenn Dahyot.
This research was supported by funding through the Maynooth University Hume Doctoral Awards.
We would like to thank the authors of the FMO datasets for making their datasets available.

License:

This code is available under the MIT License.