Cross-City Analysis

Visualizations of how SSL backbones generalize across different cities. Data is sourced live from the experiment database.

Cross-City Transfer Matrix

Best experiment per training city, showing PDM scores on each evaluation city. Diagonal entries (marked) indicate in-distribution performance.

Train City	Boston	Las Vegas	Pittsburgh	Singapore	Avg
all(A3)	84.7%	91.5%	80.3%	80.2%	85.4%
boston(P1-B1)	79.7%	77.5%	61.2%	52.8%	71.0%
vegas(P1-B2)	69.1%	92.0%	56.9%	52.5%	71.7%
pittsburgh(P1-B3)	73.2%	78.7%	73.1%	51.9%	71.7%
singapore(P1-B4)	54.7%	66.1%	53.7%	62.7%	59.6%

Backbone Comparison

Average PDM scores grouped by SSL backbone architecture.

ResNet34

15 experiments

Average PDM63.1%

Best PDM85.4%

2 experiments

Average PDM46.4%

Best PDM66.9%

ijepa

10 experiments

Average PDM0.0%

Best PDM0.0%

dinov2

10 experiments

Average PDM0.0%

Best PDM0.0%

mae

10 experiments

Average PDM0.0%

Best PDM0.0%

Sub-Metric Breakdown

Detailed scoring components across completed experiments. NC = No Collision, DAC = Drivable Area Compliance, EP = Ego Progress, TTC = Time to Collision, C = Comfort.

ID	Backbone	NC	DAC	EP	TTC	C	Avg PDM
A3	ResNet34	0.979	0.929	0.861	0.969	0.942	85.4%
A3-b	ResNet34	0.975	0.855	0.857	0.967	0.910	79.2%
A4	ResNet34	0.967	0.869	0.847	0.957	0.895	78.3%
P1-B2	ResNet34	0.954	0.784	0.834	0.938	0.941	71.7%
P1-B3	ResNet34	0.952	0.809	0.847	0.937	0.904	71.7%
P1-B1	ResNet34	0.945	0.797	0.860	0.930	0.905	71.0%
L1-B2	ResNet34	0.938	0.776	0.840	0.926	0.885	69.1%
P1-B2-b	ResNet34	0.946	0.753	0.830	0.928	0.933	68.6%
A2		0.938	0.752	0.834	0.924	0.927	66.9%
P1-B1-b	ResNet34	0.929	0.748	0.853	0.914	0.844	65.0%
L1-B1	ResNet34	0.927	0.738	0.856	0.911	0.838	64.2%
P1-B4	ResNet34	0.867	0.792	0.848	0.852	0.915	59.6%
P1-B3-b	ResNet34	0.903	0.691	0.849	0.888	0.795	57.4%
L1-B3	ResNet34	0.895	0.663	0.859	0.870	0.845	54.5%
L1-B4	ResNet34	0.852	0.702	0.836	0.837	0.860	50.9%
A1		0.720	0.538	0.754	0.710	0.814	25.9%