Abstract
FC-TrackNet is a novel deep learning framework designed for high-speed, 6D object pose tracking that excels even when trained on limited synthetic data. By optimizing for fast convergence, the system achieves a consistent tracking frequency of 90.9 Hz and outperforms state-of-the-art methods in accuracy, particularly in scenarios involving severe occlusion and drastic motion.
Introduction
In robotics and augmented reality, 6D pose tracking—continuously estimating an object's 3D position and orientation—is critical. Traditional methods often rely on expensive per-frame re-detection or require massive, hand-annotated real-world datasets. FC-TrackNet addresses these challenges by utilizing a synthetic-data-driven approach that generalizes well to the real world while requiring significantly fewer training samples to reach peak performance. It specifically targets efficiency, data scarcity, and robustness against occlusion and sudden movement.
Methods
FC-TrackNet employs a "render-and-compare" strategy. The network processes two RGB-D inputs: the current observation ($O_t$) and a rendered image ($R_{t-1}$) based on the previous frame's predicted pose.
Key technical innovations include:
- Feature Decomposition: Extracted features are decomposed into 1D directional components (horizontal and vertical) for precise spatial shift encoding.
- Coordinate-Aware Attention (CAA): This mechanism focuses the model on relevant spatial features, aiding in target discrimination within cluttered environments.
- Residual Sampling: The pipeline calculates discrepancies between observed and rendered images to predict the relative pose transition ($\Delta T$).
- Robust Loss Function: Training utilizes a loss function measuring characteristic discrepancy between pixel intensities, optimized via Mean Squared Error (MSE).
Results and Evaluation
The model was evaluated on the YCB-Video and YCBInEOAT benchmarks. A core focus was the network's performance using small synthetic datasets (6k, 8k, and 10k samples) to demonstrate rapid convergence.
Key Performance Metrics
- Speed: The system maintains a consistent tracking frequency of 90.9 Hz, suitable for real-time robotic manipulation.
- Accuracy: On YCB-Video, the model achieved 90.11% ADD and 94.99% ADD-S using only 10k training samples.
- Data Efficiency: As shown in the results, FC-TrackNet reaches high accuracy levels with significantly less data than predecessors like se(3)-TrackNet.
Discussion and Conclusion
FC-TrackNet demonstrates that high-speed tracking can be achieved with minimal synthetic data through effective architectural inductive biases. By focusing on frame-to-frame transitions rather than absolute detection, it maintains temporal consistency during fast motion. While highly robust, future work could address pose ambiguity in symmetric objects and the dependency on accurate initial pose estimation.
Summary updated