FeatUp: A Model-Agnostic Framework for Features at Any Resolution
Authors: Stephanie Fu, Mark Hamilton, Laura E. Brandt, Axel Feldmann, Zhoutong Zhang, William T. Freeman
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show that Feat Up significantly outperforms other feature upsampling and image super-resolution approaches in class activation map generation, transfer learning for segmentation and depth prediction, and end-to-end training for semantic segmentation. |
| Researcher Affiliation | Collaboration | Stephanie Fu MIT Mark Hamilton MIT, Microsoft Laura Brandt MIT Axel Feldmann MIT Zhoutong Zhang Adobe Research William T. Freeman MIT, Google |
| Pseudocode | No | The paper describes the methods textually and with equations, but does not provide any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Additionally, we provide our code at: https://tinyurl.com/28h3yppa |
| Open Datasets | Yes | For semantic segmentation, we follow the experimental setting of both (Alain & Bengio, 2016; Hamilton et al., 2022) and train a linear projection to predict the coarse classes of the COCOStuff (27 classes) training dataset... Upsamplers are trained on the Image Net training set for 2,000 steps... We adopt the experimental setting of (Lu et al., 2022c;b) to show that our JBU upsampler improves end-to-end performance on ADE20K semantic segmentation using the Segformer (Xie et al., 2021) architecture. Specifically, we train Seg Former on ADE20k Zhou et al. (2019; 2017) (20,210 training and 2,000 val) for 160k steps. |
| Dataset Splits | Yes | We report m Io U and accuracy on the validation set in Table 1. Upsamplers are trained on the Image Net training set for 2,000 steps, and we compute metrics across 2,000 random images from the validation set. We train Seg Former on ADE20k Zhou et al. (2019; 2017) (20,210 training and 2,000 val) for 160k steps. |
| Hardware Specification | No | The paper mentions "CUDA kernel" and discusses "memory overhead" and "inference speed" for its implementation and baselines, but does not specify exact GPU models, CPU types, or other hardware components used for experiments. |
| Software Dependencies | No | The paper mentions "Py Torch implementation" but does not provide specific version numbers for any software dependencies or libraries used in the experiments. |
| Experiment Setup | Yes | We outline the hyperparameters used to train Feat Up in table 7. Table 7 lists specific hyperparameters such as Num Images (1, 4), Num Jitters Per Image (10, 2), Optimizer (NAdam), Learning Rate (0.001), Image Load Size (224), Training Steps (2000), and others. |