Physics-as-Inverse-Graphics: Unsupervised Physical Parameter Estimation from Video
Authors: Miguel Jaques, Michael Burke, Timothy Hospedales
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first evaluate physical parameter estimation accuracy and future video frame prediction on 4 datasets with different non-linear interactions and visual difficulty. We then demonstrate the value of our method by applying it for data-efficient learning of vision-based control of an under-actuated pendulum. |
| Researcher Affiliation | Academia | Miguel Jaques School of Informatics University of Edinburgh Edinburgh, UK m.a.m.jaques@sms.ed.ac.uk |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not include any explicit statement about releasing code or a link to a code repository for the described methodology. |
| Open Datasets | No | The paper states 'All datasets consist of 5000 sequences for training, 500 for validation, and 500 for testing.' and describes the generation of some datasets (e.g., '2-balls spring', '3-balls gravity', '2-digits spring') and data collection from 'Open AI Gym'. However, it does not provide concrete access information (e.g., a link or citation for public access) to these generated datasets. |
| Dataset Splits | Yes | All datasets consist of 5000 sequences for training, 500 for validation, and 500 for testing. |
| Hardware Specification | No | The paper does not specify the hardware used to run the experiments, such as specific GPU or CPU models. |
| Software Dependencies | No | The paper mentions software like 'RMSProp Hinton et al. (2012)' and 'Open AI Gym (Brockman et al., 2016)', but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | For all datasets we use RMSProp Hinton et al. (2012) with an initial learning rate of 3 10 4. For the balls and digits datasets we train for 500 epochs with α = 2, and divide the learning rate by 5 after 375 epochs. For the pendulum data we train for 1000 epochs using α = 3, but divide the learning rate by 5 after 500 epochs. We train using values of (L, Tpred, Text) set to (3, 7, 20), (3, 7, 20), (3, 7, 20), (4, 12, 24) and (3, 7, 20), respectively. We use a learnable ST scale parameter initialized at s = 2 in the balls datasets and s = 1 in the digits dataset. In these datasets we set θ = 0. |