Tame a Wild Camera: In-the-Wild Monocular Camera Calibration

Authors: Shengjie Zhu, Abhinav Kumar, Masa Hu, Xiaoming Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show the effectiveness of our method through superior performance on synthetic and zero-shot testing datasets. Beyond calibration, we demonstrate downstream applications in image manipulation detection & restoration, uncalibrated two-view pose estimation, and 3D sensing. Codes/models are held here. 4 Experiments Implementation Details Our network is trained using the Adam optimizer [32] with a batch size of 8. The learning rate is 1e 5, and the training process runs for 20, 000 steps. Our RANSAC algorithm is executed on a GPU, performing Kr = 2, 048 iterations, with each iteration including a random sample of Kc = 20, 000 incidence vectors.
Researcher Affiliation Academia Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu Department of Computer Science and Engineering, Michigan State University, East Lansing, MI, 48824 {zhusheng, kumarab6, huynshen}@msu.edu, liuxm@cse.msu.edu
Pseudocode No The paper does not include a dedicated pseudocode or algorithm block.
Open Source Code No The paper states "Codes/models are held here." in the abstract, but does not provide a concrete link or explicit statement that the code for the described method is publicly released.
Open Datasets Yes In Tab. 2, we incorporate datasets of different application scenarios, including indoor, outdoor scenes, driving, and object-centric scenes. Many of the datasets utilize only a single type of camera for data collection, resulting in a scarcity of intrinsic variations. Similar to [38], we employ random resizing and cropping to synthesize more intrinsic, marked in Tab. 2 column Syn.. ... Nu Scenes [10] Calibrated Driving ... KITTI [22] Calibrated Driving ... Cityscapes [14] Calibrated Driving ... NYUv2 [53] Calibrated Indoor ... ARKit Scenes [8] Calibrated Indoor ... SUN3D [65] Calibrated Indoor ... MVImg Net [68] Sf M Object ... Objectron [1] Sf M Object ... Mega Depth [41] Sf M Outdoor ... Waymo [58] Calibrated Driving ... RGBD [55] Pre-defined Indoor ... Scan Net [16] Calibrated Indoor ... MVS [21] Pre-defined Indoor ... Scenes11 [11] Pre-defined Synthetic
Dataset Splits No The paper mentions training and testing, and augmentation strategies, but does not provide specific details on validation dataset splits (e.g., percentages, sample counts, or explicit validation sets).
Hardware Specification Yes Tested on an RTX-2080 Ti GPU, the combined network inference and calibration algorithm runs on average in 87 ms per image.
Software Dependencies No The paper mentions "Our network is trained using the Adam optimizer [32]" but does not specify versions for other software dependencies like programming languages or libraries (e.g., Python, PyTorch).
Experiment Setup Yes Implementation Details Our network is trained using the Adam optimizer [32] with a batch size of 8. The learning rate is 1e 5, and the training process runs for 20, 000 steps. Our RANSAC algorithm is executed on a GPU, performing Kr = 2, 048 iterations, with each iteration including a random sample of Kc = 20, 000 incidence vectors. During both training and testing, the image is resized to a resolution of 480 640. The threshold for determining RANSAC inliers is kx = ky = 0.02.