Introducing Probabilistic Bézier Curves for N-Step Sequence Prediction
Authors: Ronny Hug, Wolfgang Hübner, Michael Arens10162-10169
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | As an initial proof of concept, the model performance is compared to an LSTM-MDN model and recurrent Gaussian processes on two real world use-cases, trajectory prediction and motion capture sequence prediction. 5 Real world evaluation In order to provide an initial proof of concept, the performance of the proposed N-Curve MDN is compared to different state-of-the-art sequence prediction models. Further, a qualitative evaluation is presented, inspecting results of the quantitative evaluation and different aspects of the model as partly shown in previous sections. |
| Researcher Affiliation | Academia | Ronny Hug, Wolfgang H ubner, Michael Arens Fraunhofer Institute of Optronics, System Technologies, and Image Exploitation (IOSB) Gutleuthausstr. 1, 76275 Ettlingen, Germany {ronny.hug, wolfgang.huebner, michael.arens}@iosb.fraunhofer.de Fraunhofer IOSB is a member of the Fraunhofer Center for Machine Learning. |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not include a statement or link indicating that the authors' own source code for the described methodology is publicly available. It mentions a GitHub link for a comparison model from a different paper: 'with the code provided by the authors4, only walking motion is considered. The test set is modified to only contain the first 70 points of each sequence, in order to conform with the fixed sequence length currently necessary for the N-Curve MDN. Further, the data is standardized with zero mean and unitary standard deviation. In contrast to the previous experiment, a control input is given in terms of the y coordinate of the left toes for each time step (during observation and prediction). The observation and prediction lengths are set to m = 20 and n = 50 time steps. The N-Curve MDN is compared to a simple multilayer perceptron (MLP) and the recurrent 4https://github.com/zhenwendai/RGP' |
| Open Datasets | Yes | a subset of the Stanford Drone Dataset (Robicquet et al. 2016, SDD) is used, as it contains geometrically constrained scenes including e.g. junctions. For providing a higher dimensional example, sequences consisting of 59dimensional skeleton description vectors from the CMU motion capture database3 are used. 3http://mocap.cs.cmu.edu/ |
| Dataset Splits | Yes | Here, training is performed on sequences 1 to 4 from subject 35, and testing is performed on sequences 5 to 8 from the same subject. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers for its implementation. It mentions common frameworks like 'LSTM' and 'MDN' but without versioning details for the authors' code. |
| Experiment Setup | Yes | Observation and prediction lengths are fixed to m = 20 (3.2s) and n = 40 (6.6s). An N-Curve MDN with k = 3 components and 6 control points per component is used. For the FDE, the endpoint of the N-Curve with the highest weight argmaxk πk is used. Here, the number of components of the N-Curve MDN is set to k = 1 using 10 control points. For the MLP, a single hidden layer with 1000 units and tanh activation is used. In case of the RGP model, a 2 hidden layer model with 200 inducing points is used according to the evaluation performed in (Mattos et al. 2015). The sampling rate is set to 6 Hz. Further, positional information is normalized by scaling (x, y)-coordinates to conform x, y [ 1, 1]. The data is standardized with zero mean and unitary standard deviation. |