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].
Hybrid Knowledge Routed Modules for Large-scale Object Detection
Authors: ChenHan Jiang, Hang Xu, Xiaodan Liang, Liang Lin
NeurIPS 2018 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on large-scale object detection benchmarks show HKRM obtains around 34.5% improvement on Visual Genome (1000 categories) and 30.4% on ADE in terms of m AP. |
| Researcher Affiliation | Collaboration | Chenhan Jiang Sun Yat-Sen University EMAIL Hang Xu Huawei Noah s Ark Lab EMAIL Xiaodan Liang School of Intelligent Systems Engineering Sun Yat-Sen University EMAIL Liang Lin Sun Yat-Sen University EMAIL |
| Pseudocode | No | The paper describes the methodology and its components with mathematical formulas and diagrams, but it does not include a dedicated pseudocode block or algorithm description. |
| Open Source Code | Yes | Codes and trained model can be found in https://github.com/chanyn/HKRM. |
| Open Datasets | Yes | We conduct experiments on large-scale object detection benchmarks with a large number of classes: that is, Visual Genome (VG) [23] and ADE [56]. |
| Dataset Splits | Yes | We split the remaining 92960 images with objects on these class sets into 87960 and 5,000 for training and testing, respectively. In term of ADE dataset, we use 20,197 images for training and 1,000 images for testing, following [6]...MSCOCO 2017 contains 118k images for training, 5k for evaluation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper states, 'We implement all models in Pytorch [40]', but does not specify a version number for PyTorch or any other software dependency. |
| Experiment Setup | Yes | During training, we augment with ๏ฌipped images and multi-scaling (pixel size={400, 500, 600, 700, 800}). During testing, pixel size= 600 is used... We apply stochastic gradient descent with momentum to optimize all models. The initial learning rate is 0.01, reduce three times ( 0.01) during ๏ฌne-tuning; 10 4 as weight decay; 0.9 as momentum. For both VG and ADE dataset, we train 28 epochs with mini-batch size of 2 for both the baseline Faster RCNN. |