Empowering CAM-Based Methods with Capability to Generate Fine-Grained and High-Faithfulness Explanations
Authors: Changqing Qiu, Fusheng Jin, Yining Zhang
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiment In this section, we compare FG-CAM with other methods through qualitative and quantitative experiments. |
| Researcher Affiliation | Academia | Changqing Qiu1, Fusheng Jin1*, Yining Zhang2 1Beijing Institute of Technology 2Peking University changqing qiu@bit.edu.cn, jfs21cn@bit.edu.cn, zhangyining@stu.pku.edu.cn |
| Pseudocode | No | The paper provides mathematical equations describing the FG-CAM calculation steps but does not include a structured pseudocode or algorithm block. |
| Open Source Code | Yes | Our code is available at https://github.com/dongmo-qcq/FG-CAM. |
| Open Datasets | Yes | ILSVRC2012 val (2015) is used as the data set. |
| Dataset Splits | No | The paper states that a pre-trained VGG16 model is used and that 'ILSVRC2012 val' is used as the dataset for evaluation, with 'randomly selected 2000 images' for specific evaluations. However, it does not provide details on training/validation/test splits for their own experimental setup, beyond using the 'val' split for evaluation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running experiments. |
| Software Dependencies | No | The paper mentions using 'pre-trained VGG16 (...) with batch normalization from PyTorch model zoo' but does not specify software versions for PyTorch or other dependencies. |
| Experiment Setup | Yes | For each image, we resize it to (224 224 3), convert it to range [0, 1], then normalize it using mean vector [0.485, 0.456, 0.406] and standard deviation vector [0.229, 0.224, 0.225], and no further pre-processing beyond that. (...) For FG-CAM, LRP and its variants, we use the zβ rule in input layer and the z+ rule in other layers. The blurred image is realized by Gaussian blur, that is, e I = gaussian blur2d(I, ksize, sigma), where I is the original image, e I is the blurred image. In this paper, we set ksize = 51, sigma = 50. |