On-Road Object Importance Estimation: A New Dataset and A Model with Multi-Fold Top-Down Guidance

This paper addresses the problem of on-road object importance estimation, which utilizes video sequences captured from the driver's perspective as the input. Although this problem is significant for safer and smarter driving systems, the exploration of this problem remains limited. On one hand, publicly-available large-scale datasets are scarce in the community. To address this dilemma, this paper contributes a new large-scale dataset named Traffic Object Importance (TOI). On the other hand, existing methods often only consider either bottom-up feature or single-fold guidance, leading to limitations in handling highly dynamic and diverse traffic scenarios. Different from existing methods, this paper proposes a model that integrates multi-fold top-down guidance with the bottom-up feature. Specifically, three kinds of top-down guidance factors (ie, driver intention, semantic context, and traffic rule) are integrated into our model. These factors are important for object importance estimation, but none of the existing methods simultaneously consider them. To our knowledge, this paper proposes the first on-road object importance estimation model that fuses multi-fold top-down guidance factors with bottom-up feature. Extensive experiments demonstrate that our model outperforms state-of-the-art methods by large margins, achieving 23.1% Average Precision (AP) improvement compared with the recently proposed model (ie, Goal).

Paper

Similar papers

Peer review

Reviewer mvxo7/10 · confidence 3/52024-07-08

Summary

1.This paper this contributes a new large-scale dataset named Traffic Object Importance (TOI) to addresses the problem of on-road object importance estimation, which utilizes video sequences captured from the driver’s perspective as the input. 2.The author also proposes a model that integrates multi-fold top-down guidance with the bottom-up feature.

Strengths

1.This paper describes in great detail the specialized methodology and the structure of the models. 2.The scarcity of large-scale publicly available datasets hinder the development of on-road object importance estimation. 3. This paper considers the effect of traffic rule on object importance and successfully models this abstract concept by proposing an adaptive object-lane interaction mechanism.

Weaknesses

1.In page 3 the author mentions that the traffic rule is crucial for object importance and focus on the traffic line rules , but the influence of traffic rules is varied, such as signalization. Therefore, in page 4 of table 1, the author is able to provide statistics on the scenario categories of TOI dataset and the traffic rule constraints within the dataset in experiment. 2.In page 6, the author uses three common intention behaviors in driving to reflect the driver intention (i.e., turning left, going straight, and turning right). Since the video clip length is set at 16 frames, it is important to clarify if each of the three intentions corresponds to individual frames with the 16-frame clip cut during the training and testing phases, or if multiple intentions are present within the 16 frames. The authors should further elaborate and provide the proportion of each intention in the dataset. 3.Insufficient evaluation of indicators in the experimental section. The author may add another evaluation indicator. 4.The section three can include a schematic diagram of the annotation process for the dataset.

Questions

1.I consider whether 16 frames constitute interval sampling or continuous sampling, and how many types of intentional behaviors can be expressed using 16 frames in the paper. 2.The author may can add another evaluation metric for the experiment.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes.

Reviewer Yrjv6/10 · confidence 4/52024-07-12

Summary

This paper collects a new large-scale dataset and proposes a novel method that integrates multi-fold top-down guidance with the bottom feature to address the problem of on-road object importance estimation. Specifically, the dataset is almost three times larger than the current publicly dataset for on-road object importance. In addition, this paper considers an adaptive mechanism for object-lane interaction, effectively modeling the impact of traffic rules on object importance. Experiments on several benchmarks validate the effectiveness of the proposed method.

Strengths

This paper makes several key contributions and demonstrates strengths for on-road object importance estimation (1) This paper introduces a novel, extensive dataset, set to be released to the public, which is nearly three times the size of the current largest public dataset. (2) The method is well-motivated and straightforward. It estimates the importance of objects on the road, integrating various top-down guidance factors with bottom-up features, marking the first of its kind. (3) The proposed method addresses the pivotal role of traffic rules in estimating object importance, an aspect previously overlooked by existing methods. It successfully encapsulates this concept through an innovative, adaptive mechanism for object-lane interaction.

Weaknesses

This paper has also two weaknesses: (1) The paper does not provide a detailed discussion on the computational efficiency of the proposed method, which is crucial for real driving scenarios. Moreover, it is recommended to compare the model parameters and latency with other methods. (2) Another concern lies in the practicality of the method. This method and the proposed dataset are both for single-camera scenarios, but in real autonomous driving scenarios, surrounding view is a more widely used type and a safer option. Will the proposed method also work well in the surrounding view?

Questions

(1) Can the proposed method be applied to surrounding view images? I suggest that the authors should consider the application on the current perception pipeline for vision-based autonomous driving pipeline. (2) I suggest that the authors should analyze the latency of the proposed method, which determines whether the method can be integrated into the practical driving scenarios.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The proposed method only considers the effect of three types of driver intentions on object importance estimation, which is not sufficient for complex driving scenarios. I carefully checked the paper and found no potential negative societal impact.

Reviewer 8N7U6/10 · confidence 3/52024-07-16

Summary

This paper presents a novel dataset for on-road object importance estimation. More data about which objects are important for self-driving is included and is promised to be released. Moreover, a novel method that integrates driven intention, semantic context, and traffic rule is devised to tackle the related problem. The paper is well-written.

Strengths

A new dataset is introduced with rich data and labels. The presented method is novel and shown to be effective for the studied problem. Details about the dataset and the method are comprehensive and technically sound. Results are also promising.

Weaknesses

Some of the concepts lack sufficient details to explain. See questions below.

Questions

(1) Regarding the task, my major concern is the definition of importance. It is shown that surrounding objects that follow the traffic rules are not considered as important. Only the objects ahead of the car or have an intersection with the ego-car's direction are important. I doubt whether this is strictly appropriate. For example, if a pedestrian walking along the road, he/she will not be considered as important. However, what if this pedestrian suddenly steps into the road ahead, potential collisions would happen. Therefore, I think a nearby walking pedestrian should be considered as important or at least recognized into a third category like "needs care". I wonder how the authors solve this problem in the dataset. (2) Regarding the driver's intention, it is indeed difficult to define appropriately. The authors have mentioned this in the paper, but the strategy introduced to accommodate this is still not clear to me. The authors mentioned learning the intention values based on driving behaviors, but how do we know the driving behaviors? Are these behaviors (e.g. turning left) already provided in the dataset? (3) More visualization about the labels and method comparisons are better to be presented for more clarity.

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have mentioned limitations.

Reviewer xrYx6/10 · confidence 3/52024-07-22

Summary

This work addresses the issue of estimating the importance of on-road objects using video sequences from a driver’s perspective, a critical task for enhancing driving safety. The authors introduce the Traffic Object Importance (TOI) dataset, which is significantly larger and more diverse than existing datasets, and propose a novel model that integrates multi-fold top-down guidance factors—driver intention, semantic context, and traffic rules—with bottom-up features for more accurate importance estimation. Experimental results demonstrate that the proposed model significantly outperforms state-of-the-art methods in on-road object importance estimation.

Strengths

1. The introduction of the Traffic Object Importance (TOI) dataset, which is significantly larger and more diverse than existing datasets, provides a robust foundation for training and evaluating models in on-road object importance estimation, thereby addressing a major limitation in the field. 2. The proposed model effectively integrates multi-fold top-down guidance factors—driver intention, semantic context, and traffic rules—with bottom-up features, which showed good performance for the TOI task.

Weaknesses

1. Lack of description of the annotation details. How many annotators are involved in the annotation procedure? It would be good if the authors can provide some annotation procedure samples regarding the double-checking annotation mechanism and the triple-discussing annotation mechanism. 2. It seems this annotation will be varied according to different traffic rules. Since KITTI is collected in Germany, the annotators should be familiar to germany traffic rules. However the authors did not mention this information in their submission, thereby the label quality is doubtful. 3. The authors are encouraged to build up the first benchmark based on the proposed dataset by using various existing object detection methods, e.g., Yolo, with the proposed head or simpler head. It is interesting to see how the existing object detectors work on this new task. 4. More statistics of the dataset are encouraged to be given, e.g., the number of important object of different categories, etc.

Questions

1. How many annotators were involved in the annotation procedure for the dataset? Can the authors provide detailed examples of their double-checking and triple-discussing annotation mechanisms? 2. Were the annotators familiar with German traffic rules, given that the dataset was collected in Germany (KITTI dataset)? How was the expertise of the annotators in relation to German traffic laws ensured and validated? 3. Have the authors considered building the first benchmark using their dataset with existing object detection methods, such as YOLO? What were the performance outcomes of these existing methods when applied to the new task? 4. Can the authors provide more detailed statistics about the dataset, such as the number of important objects in different categories? How do these statistics compare to other datasets in the same domain?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

yes the authors mentioned it in the appendix

Reviewer xrYx2024-08-09

Response to the authors

Thank you for your response. My concerns are mostly solved. I would lile to improve my score to 6.

Reviewer 8N7U2024-08-09

In fact, I think I do not get too much new information from the authors' response to my question 1, though this would not affect my original rating. Overall this is an interesting paper. I only suggest that the authors present more discussions on the annotation process, not only on how annotations are obtained but also on the strengths and weaknesses of the applied annotation policy. Regarding the answer to my question 2, I can get the idea and I also suggest adding more related discussions in the paper.

Reviewer Yrjv2024-08-13

Post rebuttal

Thanks for the rebuttal. Most of my concerns are well-addressed therefore I tend to keep my positive rating.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC