Learning Affinity with Hyperbolic Representation for Spatial Propagation
Authors: Jin-Hwi Park, Jaesung Choe, Inhwan Bae, Hae-Gon Jeon
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a variety of experiments on spatial propagation tasks, including depth completion (Sec. 5.1) and semantic segmentation (Sec. 5.2) as shown in Fig. 4. Moreover, we provide ablation studies to describe the effects of each component in HAM and the robustness of our method concerning input sparsity and feature compression (Sec. 5.3). |
| Researcher Affiliation | Academia | 1AI Graduate School, GIST, South Korea 2KAIST, South Korea. |
| Pseudocode | Yes | Algorithm 1 Hyperbolic Affinity Learning Module |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code or links to a code repository. |
| Open Datasets | Yes | We conduct our experiment on the NYUv2 dataset (Silberman et al., 2012)... Virtual-KITTIv2 (Cabon et al., 2020) is a photo-realistic synthetic dataset... and Scan Net (Dai et al., 2017)... Augmented Pascal VOC 2012 (Everingham et al., 2015) dataset... NYUv2 (Silberman et al., 2012), Pascal-Context (Gupta et al., 2013), and ADE20K (Zhou et al., 2017)... COCO-Stuff10K (Caesar et al., 2018). |
| Dataset Splits | Yes | Using an official train/test split, we generate random depth samples as proposed in the baseline model, as n Fig. 5. We split the virtual-KITTIv2 into one scene (Scene-02) for validation and the other scenes (Scene01,06,18,20) for training without any temporal overlap between them. For Scan Net, we follow the official train/test split: 1,513 scenes for training and 100 scenes for the test. Augmented Pascal VOC 2012 (Everingham et al., 2015) dataset provides 10,582 training, 1,449 validation, and 1,456 test images with pixel-level labels in 20 foreground object classes and one background class. |
| Hardware Specification | Yes | it takes about 1 day for training networks using 4 NVIDIA RTX 3090 GPU. |
| Software Dependencies | No | We optimize our methods using Adam optimizer (Kingma & Ba, 2015) with β1 = 0.9, β2 = 0.999 with the initial learning rate of 0.001 |
| Experiment Setup | Yes | We set the curvature κ to 0.1 and a size of the kernel γ to 3, and utilize a 2-dimensional hyperbolic convolution operation for spatial propagation tasks. We optimize our methods using Adam optimizer (Kingma & Ba, 2015) with β1 = 0.9, β2 = 0.999 with the initial learning rate of 0.001, and it takes about 1 day for training networks using 4 NVIDIA RTX 3090 GPU. Our HAM is trained for 30 epochs with both L1 and L2 losses, and the initial learning rate is decayed by 0.5 at every 5 epochs after the first 10 epochs. In the training phase, we choose a batch size of 12. |