Leveraging Driver Field-of-View for Multimodal Ego-Trajectory Prediction

Dear Reviewer H88J,

We sincerely appreciate your thoughtful review and constructive feedback. We are particularly encouraged by your recognition of our solution's theoretical soundness and the value of our contributions to the dataset and metrics. We address your questions and concerns below:

1. Failure Case Analysis

We have now added a new Section D.2 to the Appendix analyzing failure modes, with particular focus on cases where predictions significantly diverge from ground truth. We observe two primary failure patterns:

• Right turns with partial occlusions by trees or buildings, and
• Movements before/after a stop sign, due to individual driver behavior patterns

These cases, particularly the stop sign scenario, highlight an interesting future direction: incorporating driver behavior modeling into trajectory prediction. We've included visualizations of these cases in the new Figure 17, along with a discussion of potential mitigation strategies.

2. Comparison with Recent Work

Thank you for this suggestion. We have extended our related work to include recent works (Lines 102-107, Lines 127-132) and summarize the changes here:

• Other recent related works we found for maneuver predictions, based on inferring driver’s intent: CEMFormer, employing in-cabin footage as a source of driver behavior; and HeadMon, using inertial head motion sensor data as intent-related data. We note that our task, trajectory prediction, is more advanced than maneuver prediction and that we focus on exploring the new driver’s field of view modality in the form of visual features to improve the prediction accuracy.
• An overview of many other recent works that makes use of different modalities.

The recent advances in trajectory prediction have primarily focused on multi-agent scenarios. In contrast, our work addresses the distinct challenge of ego-trajectory prediction using driver-centric data available in GEM and DR(eye)VE. Even the most recent work in this space, such as Ma et al.'s CEMFormer, uses in-cabin footage for intent classification rather than continuous trajectory prediction.

Our attempts to adapt state-of-the-art vehicle trajectory predictors like Autobots to our setting (by providing only ego-trajectory while zero-padding other inputs such as map data) resulted in significantly worse performance (10.87m ADE on GEM), reinforcing that our modality-specific approach and carefully chosen baselines remain most appropriate in 2024 for evaluating driver-aware ego-trajectory prediction.

3. Handling Rapid Gaze Movements and Measurement Errors

Human gaze naturally alternates between fixations (where gaze remains relatively stable on a point of interest) and saccades (rapid jumps between fixations). This fundamental pattern of human visual attention is well-documented in driving research (Land & Lee, 1994; Hayhoe et al., 2003). Following established practices in gaze processing (Pupil Labs, 2024), we handle these natural gaze patterns through:

• Standard fixation detection as described in Appendix C.1, which identifies periods where gaze remains within a small spatial region (typically 80ms to 1 second), filtering out saccades and blinks, and
• Temporal alignment through median downsampling, matching our high-frequency gaze data (200 Hz) to video timestamps (5 Hz) while ensuring robust position estimates.

Importantly, our "field-of-view" approach treats gaze as supplementary to the driver's head-mounted video feed, providing robustness to the inherent discreteness of natural gaze patterns. As such natural patterns are already present in our training data, based on the results in Table 2, they are well-handled by our methodology.

To quantify the robustness of our model to hardware-sourced gaze measurement errors, we conducted additional experiments on GEM, reporting the results in the table below. When we artificially introduce gaze position noise of ±50 pixels (~5% error), prediction accuracy only degrades by 0.01m ADE (0.02m on ADE+20PCI), demonstrating that our processing pipeline, built on standard gaze analysis techniques, effectively handles both natural gaze patterns and potential measurement uncertainties. We add this table to the Appendix in Section C.2 (the top and bottom rows were quoted from Table 2).

4. Gaze Position Representation

To clarify Line 190: Yes, we use one gaze position per timestamp. We downsample the approximately 40 gaze samples per video frame to one tuple per frame using fixation detection and median filtering. The Tx2 shape represents this final aligned data format, where each gaze position is represented as (x,y) coordinates matched to the corresponding video frame.

We appreciate your feedback and believe addressing these points has helped clarify important aspects of our work.

Dear Reviewer 15rW,

We appreciate your thorough and insightful review. Your recognition of our paper's technical depth and clarity is particularly meaningful, as we invested significant effort in presenting the complex multimodal architecture clearly. We are encouraged by your positive assessment of our key contributions - the GEM dataset, RouteFormer architecture, and PCI metric. We address your questions below.

1. PCI Definition and Bias

We appreciate your important question about potential bias in identifying corner cases, for which we use the novel PCI metric as an indicator.

As a metric of complexity, PCI compares the target ground truth trajectory $T_{target}$ with an imaginary linear trajectory $T_{simple}$, where the vehicle maintains the final speed and angle from the input sequence. A newly added Figure 10 to Appendix A.2 illustrates this "simple" trajectory in an example.

Consequently, the model predictions themselves are not used to evaluate trajectories, nor as part of a contrastive loss scheme. Rather, we use a basic extrapolation (which is also our linear baseline in Table 2) to categorize all trajectories. Through its simplicity and consistency, this design avoids introducing model bias to the evaluation phase.

PCI is crucial for our analysis: in Figures 6 and 7 it reveals that gaze's impact is very significant in the hard samples. Similarly, in Table 2, it allows us to easily filter trivial trajectories with <20 PCI, emphasizing improvements achieved via different modalities. Finally, it correlates with our expectations of long-tail trajectories, as we explore below.

2. Long-Tail Phenomenon

Consistently with both GEM and DR(eye)VE datasets, around 75% of trajectories have less than 20 PCI. Referring to Figure 4, we consider these instances as cases of straight cruising with little to no speed change. However, you rightly ask about the remaining 25% of samples.

While driver intent prediction often relies on hand-labeled events (as in the LOKI dataset from Girase et al.), with turns, lane changes and "cut-ins" requiring human consensus on the action, we take a different approach. We consider rare scenarios not as hand-labeled fixed targets, but as trajectories with unexpected continuations given initial behavior: drastic angle or speed changes, and non-linear paths.

This seemingly simple benchmark aligns surprisingly well with driver-perceived challenges. Our newly added Figure 11 in the Appendix A.3 employs a sliding window to assess average PCI across a GEM dataset video. In this sample, the following situations register as high PCI:

• Roundabouts: highest PCIs are observed in roundabouts, with using the third exit registering higher PCI than using 1st or 2nd exits.
• Lane merges: high PCI is recorded due to a slowdown and a change of direction
• Turns: turn angle appears to be an important factor that elevates PCI, together with the turn speed

Despite its simplicity, PCI powerfully detects relevant driving events without requiring map metadata, demonstrating our model's strengths.

3. Related Work Improvements

Thank you for noting the relevant ICRA and ECAI 2024 papers. We've updated our manuscript and incorporated these citations in our related work. In addition, we expand our related work discussion (Lines 102-107, Lines 127-132) and summarize the changes here:

• The addition of human-inspired trajectory prediction works as recommended by the reviewer,
• Recent maneuver prediction works using inferred driver intent: CEMFormer, which explores the benefit of in-cabin driver behavior monitoring for maneuver prediction; and HeadMon, which uses inertial head motion data for maneuver prediction. In contrast, we perform the more complex trajectory prediction task and employ driver’s visual field of view,
• A discussion on a few recent works in the multi-agent version of trajectory prediction using different modalities.

The literature remains near-exclusively multi-agent, and our paper preserves its edge for ego-trajectory prediction together with its unique addition of driver field-of-view.

4. Additional Discussion

Thank you for this valuable suggestion. We have added a new Section D.2 to the Appendix analyzing key failure modes in the GEM dataset, particularly cases involving occluded right turns and varying driver behaviors at stop signs. These examples highlight both current limitations and promising research directions in driver behavior modeling.

5. Open-Sourcing the Project Code and Dataset

We confirm that both the complete codebase and GEM dataset will be made publicly available upon publication, including:

• Full dataset (GEM) with the ground truth
• Model implementation and training code
• Data collection and preprocessing pipeline
• PCI computation tools
• GPS correction tooling

We thank you again for your constructive feedback that has helped significantly strengthen both the clarity and technical depth of our manuscript.

Dear Reviewer 15rW,

We are grateful for your thorough review and valuable feedback, which has significantly enhanced our manuscript. Your recognition of our work's contribution to the AD community is particularly meaningful, and we look forward to supporting future research through our open-sourced code and dataset.

Sincerely,

The Authors

Dear Reviewer U1qP,

Thank you for your thoughtful review and valuable feedback on our paper. We appreciate your recognition of our contributions, and have conducted additional analyses to address your concerns comprehensively, which have led us to interesting findings.

1. Ablation Studies on Modality Contributions

Following our architecture in Figure 2(a), we conducted ablation experiments by selectively removing modality branches from RouteFormer (full, trained model) during inference. For FOV removal, we bypass the cross-modal transformer and use only scene features for self-attention. Similarly, for scene removal, we retain only the FOV branch. To compare the motion-only performance with them, we replace the entire visual feature tensor $v_{1:T}$ with zeros. The results reveal several interesting findings about RouteFormer's behavior and robustness.

The results demonstrate clear benefits of each modality: removing FOV increases ADE by 10.2%, while removing scene information leads to a 30.7% increase in error.

Notably, our scene-only variant maintains performance consistent with the RouteFormer-Base (6.13m), which was trained directly with these inputs, and our FOV-only variant remains competitive with GIMO (7.61m) as shown in Table 2. In addition, the motion-only variant still shows significant improvement over the linear baseline (13.37m) for ADE+20PCI range. This stability across modality configurations is particularly valuable for real-world deployment, where sensors or gaze tracking might be temporarily unavailable, allowing graceful performance degradation rather than catastrophic failure when inputs are missing.

These findings have been added to a new ablation studies section of the Appendix D.3.

2. Choice of Baseline Methods

We appreciate the suggestion to include vehicle-specific models like MTR [1] and Autobots [2] as baselines to strengthen our evaluation. We have carefully considered this and would like to provide a detailed explanation of our approach.

We attempted to adapt Autobots to our datasets by using only the available inputs, and train it on GEM from scratch. Specifically, we provided the past ego-trajectory (ego_in) while setting trajectory of other agents (agents_in) and road graph information (roads) to zeros or placeholders, as expected by the model signature. However, the model's performance was significantly worse than our baselines, yielding an ADE of 10.87m on GEM.

This is due to the native incompatibility of our task with the input modalities, as explained in the table below.

Overall, multi-agent trajectory prediction models like Autobots and MTR are architecturally designed to leverage rich environmental context from multi-agent interactions. Ego-trajectory prediction, however, is an inherently different task that relies on the view and intention of the driver, often without the availability of additional metadata.

Given these constraints, we focused on baselines that enable fair evaluation of our core contribution - integrating driver FOV data for trajectory prediction. GIMO and Multimodal Transformer were selected as they support similar input modalities while representing SOTA in attention-based trajectory prediction.

In light of this question, we expanded our discussion of baseline selection criteria with a section D.4 in the Appendix, and include detailed comparisons of architectural differences between ego-motion and multi-agent prediction approaches. We thank you again for your constructive feedback that has helped strengthen our experimental validation.