Phase 2: Base Projects (Which project should we base on?)

[wangyashuu]

Base Project Research

SocialGAN

Dataset

from paper - Section 4. Experiment

In this section, we evaluate our method on two publicly available datasets: ETH [36] and UCY [25]. These datasets consist of real world human trajectories with rich humanhuman interaction scenarios. We convert all the data to real world coordinates and interpolate to obtain values at every 0.4 seconds. In total there are 5 sets of data (ETH - 2, UCY3) with 4 different scenes which consists of 1536 pedestrians in crowded settings with challenging scenarios like group behavior, people crossing each other, collision avoidance and groups forming and dispersing.

Result

Code to Read

• about data
• about model
• about train

Running Example

Notebook

• running time: about 7 hours? (half number epochs cost 3.5 hours)
• unknown, timeout.

Issues

• They convert image coordinates to real world coordinates for UCY dataset. But this preprocess is private. ref issue

SocialWays

Dataset

from paper - Section 4.3 Datasets

For the evaluation of our approach, we use two publicly available datasets: ETH [13] and UCY [11]. These datasets consist of real-world human trajectories. They are labeled manually at a rate of 2.5 fps. The ETH dataset contains 2 experiments (coined as ETH and Hotel) and the UCY dataset contains 3 experiments (ZARA01, ZARA02 and Univ). In order to evaluate the prediction algorithm, each dataset is split into 5 subsets, where we train and validate our model on 4 sets and test on the remaining set.

Result

Code to Read

• about data
• about train, model

Running Example

Notebook

• running time: 16min on Colab, 11min on Gradient
• running result (eth): ADE/FDE 0.039, 0.073 (ADE/FDE on Paper 0.39, 0.64)

Issues

• They do not provide parse step for ucy dataset (ref issue). The author recommends to use the dataset he uploaded on dropbox (ref comment).
• TODO: Is it wrong for infoGAN part?

Trajnet++

Dataset

Trajnetplusplusdataset

This repository provides scripts to generate train, val and test dataset splits from raw data as well as simulators.

Issues

• It did not provides a clear comparation result for baseline ref projects. But we might can run for a result for original Trajnet++.
• Seem like for ucy dataset it did not do specical handle (convert image cordinate to real world cordinate). But still the project make some other preprocess.

[wangyashuu]

See Socialways problems mentioned in Phase3GettingResults