Fahad Majeed
Portrait of Fahad Majeed

Fahad Majeed, PhDComputer vision researcher
Hamad Bin Khalifa University, Doha, Qatar

I teach machines to read a soccer match from video.

My research turns broadcast footage into things a coach or analyst can use: who is where, who is close to the ball, how far the scene is from the camera, and what the commentators are saying about it. I have spent more than ten years across academia and industry building these systems end to end.

Illustration of the ball and player graph behind GameFlow. Yellow edges link the ball to its nearest players in each frame.

Current work

The questions I am spending most of my time on now.

  • Depth from a single broadcast camera

    General depth models struggle on a pitch: fast motion, repeated grass texture and heavy occlusion. I work on fusing several depth experts so the right one is trusted at each pixel.

  • Listening to the match as well as watching it

    Commentary carries context that video alone misses. I build pipelines that clean, segment, transcribe and translate match audio so it can be paired with what happens on screen.

  • Graph reasoning for player behaviour

    Players and the ball form a graph that changes every frame. I use graph neural networks over these sequences to describe interactions, movement patterns and events.

  • Vision beyond the pitch

    The same tools carry over to other domains. I have applied detection and segmentation models to medical imaging, including 3D segmentation on BraTS and ATLAS, and to surveillance video.

Research projects

Five published projects, newest first. Each one covers a different piece of the same problem: understanding a soccer match from broadcast video and audio.

IEEE Open Journal of Signal Processing 2026, presented at ICIP 2026
Depth estimation

DepthPlay-Agent (DEAN)

Agentic learning and fusion for depth estimation

No single depth model is best everywhere in a sports scene. DEAN treats four strong depth estimators as experts and learns a mixture-of-experts controller that decides, pixel by pixel, which of them to rely on. A domain-aware refinement step then cleans up the fused map.

  • RGB frame
  • UNet++, HybridDepth, DINOv3, ZoeDepth
  • Mixture-of-experts controller
  • Domain-aware refinement
  • Depth map

Evaluated on SoccerNet-Depth, KITTI, MPI-Sintel and NYU Depth v2, so the method is tested on sports footage as well as driving, synthetic and indoor scenes.

Scientific Reports 2026, Nature Portfolio
Audio and language

EchoNet++

A multilingual soccer match audio commentary dataset

A dataset and a reproducible pipeline for broadcast commentary. The pipeline isolates speech, removes crowd noise, separates commentator speech from spectator sound, transcribes it with several Whisper variants and translates non-English commentary into English. The paper also measures how much each stage contributes.

  • Match video
  • Band-pass filter and Demucs denoising
  • Silero voice activity detection
  • Whisper transcription
  • Translation
  • Time-stamped JSON per match

Tested on full matches from Europe's top leagues under real broadcast conditions.

Scientific Reports 2025, Nature Portfolio
Graph learning

GameFlow

Real-time analysis of soccer ball and player interactions using graph convolutional networks

GameFlow keeps detection and graph reasoning as separate stages so the whole system can run in real time. A CSPDarknet53 detector finds players and the ball, a graph is built for every frame from distances between them, and a GCN reads the graph sequence to recognise interactions and events. It also reports distance covered and speed for each player. The work introduces the Soccer++ dataset.

  • Video frames
  • Detection and tracking
  • Ball and player graph per frame
  • Graph convolutional network
  • Interactions, events, speed
Object detection
91%
Tracking and action recognition
90%
Speed analysis
92%
VISAPP 2025 Porto, Portugal
Tracking and segmentation

ReST

High-precision soccer player tracking via motion vector segmentation

ReST stands for recognition, segmentation and tracking. It feeds the network motion vectors, computed with a Scharr filter and frame differencing, next to the RGB frame. The extra shape and motion cues help keep players apart when they overlap or crowd together. The backbone is GELAN, and the model is trained on several datasets at once to reduce dataset bias.

  • RGB frames
  • Scharr edges and frame differencing
  • Motion vectors
  • GELAN backbone
  • Player masks and tracks
DFL Bundesliga Data Shootout
97%
SoccerNet-Tracking
98%
SoccerPro
99%
From left to right: RGB frame, Scharr edges of the frame, colour-coded motion vectors, and edges of the motion vectors.
CVSports at CVPR 2024 Seattle, USA
Tracking and segmentation

MV-Soccer

Motion-vector augmented instance segmentation for soccer player tracking

The project that started the motion-vector line of work. RGB frames alone say little about shape when players are small and partly hidden. MV-Soccer adds motion vectors and frame differences as extra input, which makes it easier to separate players from the background and from each other. It uses a CSPDarknet53 backbone and is trained jointly on two public datasets and SoccerPro, a private dataset built for this work.

  • RGB frames
  • Frame differencing and motion vectors
  • CSPDarknet53 backbone
  • Detection and instance segmentation
  • Player tracks
DFL Bundesliga Data Shootout
97%
SoccerNet-Tracking
98%
SoccerPro
99%

Publications

Peer-reviewed papers. A full and current list is on Google Scholar.

  1. 2026IEEE OJSP, ICIP 2026

    DepthPlay-Agent: Agentic Learning and Fusion for Depth Estimation

    Fahad Majeed, Maria Nazir, Zhihe Lu, Marco Agus, Jens Schneider

    IEEE Open Journal of Signal Processing, vol. 7, pp. 678–687. Presented at IEEE ICIP 2026.

    Paper
    BibTeX
    @article{majeed2026depthplay,
  title   = {DepthPlay-Agent: Agentic Learning and Fusion for Depth Estimation},
  author  = {Majeed, Fahad and Nazir, Maria and Lu, Zhihe and Agus, Marco and Schneider, Jens},
  journal = {IEEE Open Journal of Signal Processing},
  volume  = {7},
  pages   = {678--687},
  year    = {2026},
  doi     = {10.1109/OJSP.2026.3703724}
}
  2. 2026Scientific Reports

    EchoNet++: A Multilingual Soccer Match Audio Commentary Dataset

    Fahad Majeed, Maria Nazir, Marco Agus, Jens Schneider

    Scientific Reports, vol. 16, article 8884. Nature Portfolio.

    Paper
    BibTeX
    @article{majeed2026echonet,
  title   = {EchoNet++: A Multilingual Soccer Match Audio Commentary Dataset},
  author  = {Majeed, Fahad and Nazir, Maria and Agus, Marco and Schneider, Jens},
  journal = {Scientific Reports},
  volume  = {16},
  number  = {1},
  pages   = {8884},
  year    = {2026},
  doi     = {10.1038/s41598-026-39884-8}
}
  3. 2025Scientific Reports

    Real-time Analysis of Soccer Ball–Player Interactions Using Graph Convolutional Networks for Enhanced Game Insights

    Fahad Majeed, Maria Nazir, Kamilla Swart, Marco Agus, Jens Schneider

    Scientific Reports, vol. 15. Nature Portfolio.

    Paper
    BibTeX
    @article{majeed2025gameflow,
  title   = {Real-time analysis of soccer ball--player interactions using graph convolutional networks for enhanced game insights},
  author  = {Majeed, Fahad and Nazir, Maria and Swart, Kamilla and Agus, Marco and Schneider, Jens},
  journal = {Scientific Reports},
  volume  = {15},
  year    = {2025},
  doi     = {10.1038/s41598-025-05462-7}
}
  4. 2025VISAPP

    ReST: High-Precision Soccer Player Tracking via Motion Vector Segmentation

    Fahad Majeed, Khaled Ahmed Lutf Al Thelaya, Nauman Ullah Gilal, Kamilla Swart-Arries, Marco Agus, Jens Schneider

    20th International Conference on Computer Vision Theory and Applications (VISAPP), vol. 2, pp. 138–149. Porto, Portugal.

    Paper PDF
    BibTeX
    @inproceedings{majeed2025rest,
  title     = {ReST: High-Precision Soccer Player Tracking via Motion Vector Segmentation},
  author    = {Majeed, Fahad and Al Thelaya, Khaled Ahmed Lutf and Gilal, Nauman Ullah and Swart-Arries, Kamilla and Agus, Marco and Schneider, Jens},
  booktitle = {Proceedings of the 20th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 2: VISAPP},
  pages     = {138--149},
  year      = {2025},
  doi       = {10.5220/0013168000003912}
}
  5. 2024CVPR Workshops

    MV-Soccer: Motion-Vector Augmented Instance Segmentation for Soccer Player Tracking

    Fahad Majeed, Nauman Ullah Gilal, Khaled Al-Thelaya, Yin Yang, Marco Agus, Jens Schneider

    10th International Workshop on Computer Vision in Sports (CVSports), IEEE/CVF CVPR Workshops, pp. 3245–3255. Seattle, USA.

    Paper Code
    BibTeX
    @inproceedings{majeed2024mvsoccer,
  title     = {MV-Soccer: Motion-Vector Augmented Instance Segmentation for Soccer Player Tracking},
  author    = {Majeed, Fahad and Gilal, Nauman Ullah and Al-Thelaya, Khaled and Yang, Yin and Agus, Marco and Schneider, Jens},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  pages     = {3245--3255},
  year      = {2024},
  doi       = {10.1109/CVPRW63382.2024.00330}
}

Datasets and tools

Data I have built, data I work with, and the stack I use day to day.

Datasets I built

  • EchoNet++: multilingual soccer commentary with transcripts, translations and timestamps.
  • Soccer++: full-match soccer video introduced with GameFlow.
  • SoccerPro: private player detection, segmentation and tracking data used in MV-Soccer and ReST.

Datasets I work with

SoccerNet, SoccerNet-Tracking, SoccerNet-Depth, DFL Bundesliga Data Shootout, FIFA World Cup data, KITTI, NYU Depth v2, MPI-Sintel, BraTS and ATLAS (ISLES 2022). Custom annotation with CVAT, LabelMe, Makesense.ai and Roboflow.

Vision

Detection with the YOLO family, RF-DETR, Faster R-CNN and Mamba-YOLO. Instance segmentation with Mask R-CNN, and MedSAM2 for 3D medical images. Multi-object tracking with SORT, OC-SORT, Deep SORT, ByteTrack and BoT-SORT. Pose estimation with OpenPose, HRNet and MediaPipe.

Graphs and time

GCN, GAT, EdgeConv and ST-GCN. Spatio-temporal adjacency modelling and action prediction from pose graph sequences.

Engineering

PyTorch, TensorFlow, Detectron2, MMDetection, scikit-learn. Python, MATLAB, Bash, Docker, Git, LaTeX, REST APIs, Streamlit and Flask. Distributed video annotation with Apache Spark and cloud GPUs.

Teaching and mentoring

Alongside research I have taught undergraduate computer science and supervised student projects.

Courses taught

Artificial Intelligence, Theory of Automata, System Analysis and Design, Programming Fundamentals, Mobile Application Development, and Internet of Things.

Supervision

Undergraduate research projects and theses in computer vision, artificial intelligence, machine learning and the Internet of Things.

Industry projects

AI-driven medical diagnostics, surveillance systems and sports technology for startups and research labs.

Working on sports video, tracking or depth? Let's talk.

I am happy to hear from researchers, clubs and companies about collaborations, datasets and talks. I am based at Education City, Qatar Foundation, Doha.