🚀 Feature
I would like to request the implementation of Novel Trajectory Agent Intersection over Union (NTA-IoU) and Novel Trajectory Lane Intersection over Union (NTL-IoU) metrics (see definitions below in Additional Context).
Motivation
Both these metrics are used in autonomous vehicle research to measure the spatiotemporal accuracy of vehicles in the foreground (NTA-IoU) and geometric fidelity and spatiotemporal coherence of the road lanes in the background (NTL-IoU).
Pitch
An implementation of this exists in the DriveDreamer4D repo and can be implemented to fit the codebase here.
Alternatives
Do not know any
Additional context
(From https://arxiv.org/pdf/2511.21113)
Novel Trajectory Agent IoU (NTA-IoU): Measures the spatiotemporal accuracy of foreground dynamic agents (vehicles). Its computation involves detecting 2D bounding boxes on the rendered frames and comparing them to the projected ground-truth 3D bounding boxes. High NTA-IoU ensures accurate agent placement and adherence to the underlying 3D structure, leading to precise
corrections in under-constrained regions.
Novel Trajectory Lane IoU (NTL-IoU): Measures the geometric fidelity and spatiotemporal coherence of background lane lines. By comparing lane lines detected in the synthesized image against projected ground truth (often derived from the HDMap), it specifically verifies the integrity of the environment’s static geometry. This metric reflects minimal disturbance to the original scene structure and guarantees environmental consistency.
A few popular papers that implement this in their own codebases:
FaithFusion: https://arxiv.org/pdf/2511.21113
DriveDreamer4d: https://arxiv.org/pdf/2410.13571
🚀 Feature
I would like to request the implementation of Novel Trajectory Agent Intersection over Union (NTA-IoU) and Novel Trajectory Lane Intersection over Union (NTL-IoU) metrics (see definitions below in Additional Context).
Motivation
Both these metrics are used in autonomous vehicle research to measure the spatiotemporal accuracy of vehicles in the foreground (NTA-IoU) and geometric fidelity and spatiotemporal coherence of the road lanes in the background (NTL-IoU).
Pitch
An implementation of this exists in the DriveDreamer4D repo and can be implemented to fit the codebase here.
Alternatives
Do not know any
Additional context
(From https://arxiv.org/pdf/2511.21113)
Novel Trajectory Agent IoU (NTA-IoU): Measures the spatiotemporal accuracy of foreground dynamic agents (vehicles). Its computation involves detecting 2D bounding boxes on the rendered frames and comparing them to the projected ground-truth 3D bounding boxes. High NTA-IoU ensures accurate agent placement and adherence to the underlying 3D structure, leading to precise
corrections in under-constrained regions.
Novel Trajectory Lane IoU (NTL-IoU): Measures the geometric fidelity and spatiotemporal coherence of background lane lines. By comparing lane lines detected in the synthesized image against projected ground truth (often derived from the HDMap), it specifically verifies the integrity of the environment’s static geometry. This metric reflects minimal disturbance to the original scene structure and guarantees environmental consistency.
A few popular papers that implement this in their own codebases:
FaithFusion: https://arxiv.org/pdf/2511.21113
DriveDreamer4d: https://arxiv.org/pdf/2410.13571