Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Learning Auxiliary Monocular Contexts Helps Monocular 3D Object Detection
Authors: Xianpeng Liu, Nan Xue, Tianfu Wu1810-1818
AAAI 2022 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, the proposed Mono Con is tested in the KITTI benchmark (car, pedestrian and cyclist). |
| Researcher Affiliation | Academia | Xianpeng Liu1, Nan Xue2 , Tianfu Wu 1 1 Department of Electrical and Computer Engineering, North Carolina State University, USA 2 School of Computer Science, Wuhan University, China |
| Pseudocode | No | The paper describes the method using prose and mathematical equations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is released at https://git.io/Mono Con. |
| Open Datasets | Yes | In experiments, the proposed Mono Con is tested in the KITTI benchmark (car, pedestrian and cyclist) (Geiger et al. 2013). |
| Dataset Splits | Yes | For ablation studies, we follow the protocol used by prior works (Chen et al. 2016, 2015, 2017) to split the provided whole training data into a training subset (3,712 images) and a validation subset (3,769 images). |
| Hardware Specification | Yes | Thanks to the simple design, the proposed Mono Con obtains the fastest speed with 38.7 fps (on a single NVIDIA 2080Ti GPU card) in comparisons. |
| Software Dependencies | No | The paper mentions optimizers (Adam W) and normalization techniques (Attentive Normalization, Batch Norm) but does not provide specific version numbers for general software dependencies such as programming languages or deep learning frameworks. |
| Experiment Setup | Yes | Our Mono Con is trained on a single GPU with a batch size of 8 in an end-to-end way for 200 epochs. The Adam W optimizer is used with (Ξ²1, Ξ²2) = (0.95, 0.99) and weight decay 0.00001 (not applying to feature normalization layers and bias parameters). The initial learning rate is 2.25e 4, and the cyclic learning rate scheduler is used (1 cycle), which ο¬rst gradually increases the learning rate to 2.25e 3 with the step ratio 0.4, and then gradually drops to 2.25e 4 1.0e 4 (i.e., the target ratio is (10, 1.0e 4)). The cyclic scheduler is also applied for the momentum with the target ratio (0.85/0.95, 1) and the same step ratio 0.4. |