Fig. 1. Schematic of the PnP problem with scaling factors

Fig. 2. Pose errors in the synthetic data experiments. (a)‒(d) Rotation matrix error at different noise standard deviations;

Fig. 3. Errors in pose and scaling factors under noise standard deviation σ=5 pixel. (a) Rotation matrix error; (b) translation vector error; (c) error in scaling factor s₁; (d) error in scaling factor s₂

Fig. 4. Qualitative results of some data on the KeypointNet dataset. (a) 3D point cloud of a bottle after scaling; (d) 3D point cloud of a car after scaling; (b)(e) visualized qualitative results of our method; (c)(f) visualized qualitative results of APnP method

Fig. 5. Four images from the ETH3D Benchmark dataset, where the first two are outdoor scenes and the last two are indoor scenes

Fig. 6. Comparison of pose errors and CPU running time in real data experiments. (a)‒(d) Rotation matrix error under actual scenarios of Courtyard, Facade, Delivery area, and Statue; (e)‒(h) translation vector error under actual scenarios of Courtyard, Facade, Delivery area, and Statue; (i)‒(l) CPU running time under actual scenarios of Courtyard, Facade, Delivery area, and Statue

Fig. 7. Histograms of pose errors of each method under extreme scaling factors. (a) Rotation matrix error under the actual scenario Courtyard; (b) rotation matrix error under the actual scenario Delivery area; (c) translation vector error under the actual scenario Courtyard; (d) translation vector error under the actual scenario Delivery area