Introduction
Two obstacles motivate this paper. First, DeTra has no public implementation, making it difficult to reproduce or extend. Second, even with a documented reimplementation, forecasting dynamic (moving) actors remains markedly harder than detection: static actors dominate urban LiDAR scenes and inflate aggregate metrics, masking the moving-actor shortfall that matters most for safety. The practical goal is therefore to improve dynamic forecasting without adding inference cost or destabilizing detection.
We study complementary, low-cost inductive biases from signals not supervised by DeTra. Each actor’s observed past trajectory and the spatial occupancy of surrounding traffic are recoverable from the same annotations used to train the future loss; the agreement between an actor’s heading and its direction of motion is instead a consistency relation between the model’s own outputs. We turn each signal into a train-only auxiliary objective that shapes the shared BEV representation and is removed at inference, leaving latency unchanged. Because these objectives share the same backbone, we also calibrate their interaction: we measure each term’s gradient norm at the shared trunk and use this diagnostic to choose stable auxiliary weights and explain which signals dominate or become gradient-starved.
The key contributions of this paper are as follows:
We propose MC-DeTra, a novel training methodology that combines two annotation-derived auxiliary objectives—Past Reconstruction and Occupancy Auxiliary—and a lightweight heading-consistency regularizer studied as part of the auxiliary set, with gradient-norm loss calibration. The method reuses already-cached signals absent from DeTra’s supervision and diagnoses how auxiliary objectives compete at the shared backbone.
Motion-Consistent DeTra
MC-DeTra builds on our documented DeTra reimplementation and augments it with three train-only auxiliary objectives. We summarize the base architecture, implementation differences, objectives, and calibration.
Occupancy Auxiliary
We also implemented a novel trajectory–occupancy consistency term that would tie each decoded trajectory to its implied occupancy through differentiable soft rasterization. Gradient-norm analysis showed that this term was gradient-starved at all stable weights, so it is disabled in all reported models and treated as a negative result.
Heading Consistency
This constraint is not satisfied by degrading the forecast toward the yaw: the future waypoints are pinned by the dominant winner-takes-all Laplace regression, so HC’s small-weight gradient into the heading-consistency objective’s forecast representation is negligible. In practice, the term is met almost entirely by rotating the box yaw toward the predicted motion. We do not constrain the finite-difference future headings, which are degenerate with respect to their own waypoints.
Discussion
The transferable lesson is methodological: in a shared-backbone joint model, what matters is not conceptual importance alone but the trunk-gradient fraction an auxiliary contributes. The reported ablation pattern is consistent with this interpretation: Past Reconstruction and Occupancy Auxiliary both help, while Heading Consistency’s contribution is small and metric-specific. This reframes Heading Consistency’s apparent inertness as gradient starvation and warns against over-weighting a starved term until it distorts the head.
Conclusion
We presented an openly released DeTra reimplementation with documented approximations and MC-DeTra: three train-only, inference-safe regularizers—Past Reconstruction, Occupancy Auxiliary, and Heading Consistency—plus gradient-norm calibration of shared-backbone competition. A dynamic controller matches, but does not beat, hand-tuning. In a controlled ablation from one initializer, the calibrated full model attains the best dynamic minFDE, minADE, and brier-minFDE, with the best MR coming from controller A, while preserving detection. Gradient norms attribute the gain: occupancy dominates, past reconstruction helps modestly, and heading consistency is a small metric-dependent adjustment. We release code, configurations, and evaluation tooling for future work.
The authors have no competing interests to declare.
Frequently Asked Questions
What is MC-DeTra? MC-DeTra is a training methodology that combines Past Reconstruction, Occupancy Auxiliary, and Heading Consistency objectives with gradient-norm calibration.
Why are the auxiliary objectives removed at inference? They shape the shared bird’s-eye-view representation during training but are train-only objectives, leaving inference latency unchanged.
Why is Occupancy Auxiliary disabled in reported models? Gradient-norm analysis showed that it was gradient-starved at all stable weights.
What did gradient analysis reveal? Occupancy dominated the gain, Past Reconstruction helped modestly, and Heading Consistency provided a small metric-dependent adjustment.