Potential Field as Scene Affordance for
Behavior Change-Based
Visual Risk Object Identification

We evaluate nine baselines within our framework. These baselines take a sequence of images as input and output a risk score for each road user (e.g., vehicle or pedestrian). A road user is considered a risk object if the raw score exceeds a predefined threshold. It is important to note that all the following baselines rely on Visual-ROI methods.

• Cost Map:   FF [5].
• Collision Anticipation:   DSA [6] and RRL [7].
• Behavior Prediction-based:   BP [8], BCP [3], TP+BCP (BCP w/ Target Point) and BS+BCP (BCP w/ BEV-SEG).
• Planning-based:   OFDE and OADE.

We evaluate the performance of Visual-ROI models with two types of metrics: Spatial Accuracy and Temporal Consistency. Spatial Accuracy include the Optimal F1 Score (OT-F1) and the Optimal F1 Score in T Seconds (OT-F1-T), which measure the model's ability to accurately identify risks. Temporal Consistency metrics consist of the Progressive Increasing Cost (PIC) and the Weighted Multi-Object Tracking Accuracy (wMOTA), which assess the consistency and accuracy of the model over time.

• Optimal F1 Score (OT-F1) : An object is a risk object if its raw risk score exceeds a certain threshold. The optimal threshold is selected by maximizing the F1 score through a sweeping analysis. This process serves as an upper-bound performance benchmark for each model.


• Optimal F1 Score in T Seconds (OT-F1-T) : OT-F1-T evaluates the OT-F1 prediction outcomes during the T seconds preceding the critical point. A critical point is defined as the moment when the ego vehicle is both influenced by the risk object and is at its closest proximity to it [9].


• Progressive Increasing Cost (PIC) : This metric is introduced in RiskBench [9]. To address the issue of penalty weights decreasing too quickly and approaching zero (Fig. 1), we adjust PIC as \[ \textrm{PIC} = -\sum^{T}_{t=1} e^{-(T-t)/T}\log(\textrm{F1}_{t}) \] Here, \( \textrm{F1}_t \) denotes the F1 score at a specific time frame \( t \), while \( T \) represents the total number of frames within a scenario. We establish \( T \) as 60, equivalent to 3 seconds. We scale PIC to a range between 0 and 1 for improved interpretability by aggregating the PIC values across all scenarios and normalizing the total PIC to fit within this scale.


• Weighted Multi-Object Tracking Accuracy (wMOTA) : Inspired by MOT16 [10], we use MOTA to evaluate the temporal consistency of a Visual-ROI model. To address the imbalance between positive (risky) and negative (non-risky) samples, we propose a weighted version of MOTA called wMOTA. We denote number of positive miss at time \(t\) as \(\textrm{PM}_t\). The number of negative miss at time \(t\) as \(\textrm{NM}_t\). The value \(\textrm{PM}_t\) is defined as \(w_p\cdot(\textrm{FN}_t+\textrm{IDsw}^{p}_t)\). The value \(\textrm{NM}_t\) is defined as \(w_n\cdot(\textrm{FP}_t+\textrm{IDsw}^{n}_t)\). In the above two equations, the notations \(\textrm{FN}_t\) and \(\textrm{FP}_t\) represent the numbers of false negatives and false positives at time \(t\), respectively. In addition, the notations \(\textrm{IDsw}^p_t\) and \(\textrm{IDsw}^n_t\) represent the number of identity switches for positive and negative samples, respectively, between times \(t-1\) and \(t\). The parameters \(w_p\) and \(w_n\) are the weights assigned to positive and negative samples. The wMOTA is defined as \[ \textrm{wMOTA}=1-\frac{\sum_t(\textrm{PM}_t+\textrm{NM}_t)}{\sum_t(w_p{\cdot}\textrm{GT}^{p}_t+w_n{\cdot}\textrm{GT}^{n}_t)} \] where \(\textrm{GT}^{p}_t\) and \(\textrm{GT}^{n}_t\) represent the counts of ground truth positive and negative samples at time \(t\), respectively.