Molmo 2 | Video Tracking

2025-12-16 · Source: Ai2 · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Intermediate, medium

Summary

Momo 2 introduces enhanced video tracking capabilities, allowing users to identify and follow objects over time using point-based tracking. The system assigns persistent numerical IDs to objects based on their appearance order, maintaining these IDs even when objects are occluded or temporarily disappear and reappear, as demonstrated with penguins and cars. Unlike traditional point tracking, Momo 2 focuses on object tracking, adjusting the tracking point to the most visible part of an object during occlusion. The model leverages its video-specific architecture to access the entire video context, enabling robust re-identification of objects like car 71, which an image-only model would struggle with due to partial visibility. Users can specify tracking queries via text, such as "track the penguins" or "track the car that passes car 13," and control the sampling rate (1 or 2 FPS). An experimental UI feature also allows users to add a tracking point directly on an object and issue a simple "track" command.

Key takeaway

For AI Engineers developing video analysis solutions, Momo 2's object tracking capabilities demonstrate the critical role of video models over image-only approaches for robust re-identification and handling occlusions. You should consider integrating models with full video context access to achieve persistent object tracking, especially in dynamic scenes where objects may be partially visible or temporarily disappear, enhancing the reliability of your tracking systems.

Key insights

Momo 2's video tracking combines strong perception and reasoning for robust object re-identification across occlusions and disappearances.

Principles

• Object IDs persist through occlusion and re-appearance.
• Video models excel in tracking over image-only models.
• Tracking points adapt to object visibility.

Method

Momo 2 assigns numerical IDs to objects based on appearance order, tracks their XY coordinates, and re-identifies them using appearance and context from the full video, even under occlusion or temporary disappearance.

In practice

• Use text queries like "track the cars" for object tracking.
• Specify sampling rates (1 or 2 FPS) for tracking.
• Utilize the "add tracking point" UI feature for direct object selection.

Topics

• Video Tracking
• Object Re-identification
• Occlusion Handling
• Video Models
• Natural Language Queries

Best for: Machine Learning Engineer, Computer Vision Engineer, AI Engineer

Related on AIssential

• Molmo 2 | Counting objects and actions — Momo 2 excels in video object and action counting/pointing, while Gemini 3 shows stronger temporal understanding.
• Mastering Multi-Object Tracking with Roboflow Trackers & OpenCV — Roboflow Trackers provides modular, detector-agnostic multi-object tracking with robust algorithms like ByteTrack and flexible integration.
• Molmo 2 | Robotics Applications — Momo 2 interprets human task videos to generate actionable steps for robotics.
• SAM 3: Building a unified model architecture for detection and tracking — SAM 3 unifies object detection and tracking using distinct transformers for each task.
• MDE-VIO: Enhancing Visual-Inertial Odometry Using Learned Depth Priors — Integrating temporally consistent learned depth priors into VIO backend optimization significantly improves accuracy and robustness on edge devices.
• Multi-HMR 2: Multi-Person Camera-Centric Human Detection, Mesh Recovery and Tracking — Multi-HMR 2 integrates camera-centric detection, mesh recovery, and tracking for robust 3D human perception.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Ai2.