Notice: This website is continuously updated as new conference proceedings and journal volumes are published. Consequently, the figures presented here may differ from those reported in our publications.
A brief overview of our methodology is provided on the Methods page, along with references to publications documenting our approach and links to the code and data used in each study.
The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates.
RenderGAN: Generating Realistic Labeled Data
Authors: Leon Sixt, Benjamin Wild, Tim Landgraf
ICLR 2017 | Venue PDF | LLM Run Details | Input Tokens: 15,572 Total number of tokens sent to the LLM as input for this paper's analysis. | Output Tokens: 2,143 Total number of tokens produced by the LLM (including reasoning/thinking tokens) for this paper's analysis.
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Training a DCNN on data generated by the Render GAN yields considerably better performance than training it on various baselines. We furthermore include a previously used computer vision pipeline in the evaluation. The networks detections are used as feature to track the honeybees over time. When we use detections from the DCNN instead of the computer vision pipeline, the accuracy of assigning the true id increases from 55% to 96%. |
| Researcher Affiliation | Academia | Leon Sixt, Benjamin Wild, & Tim Landgraf Fachbereich Mathematik und Informatik Freie Universit at Berlin Berlin, Germany EMAIL |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We will outline only the most important parts. See our code available online for all the details of the networks1. 1https://github.com/berleon/deepdecoder |
| Open Datasets | No | The paper mentions using |
| Dataset Splits | Yes | As the data is correlated in time (same ids, similar lighting conditions), we assign the data from one time span completely to either the train or test set. The data from three time spans forms the train set (40K). The test set (26K) contains data from the remaining two time spans. ... For the training with generated data, we use the real training set as the validation set. When training on real data, the test set is also used for validation. We could alternatively reduce the real training set further and form an extra validation set, but this would harm the performance of the DCNN trained on the real data. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions using Adam as an optimizer but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | Training: We train generator and discriminator as in the normal GAN setting. We use 2.4M unlabeled images of tags to train the Render GAN. We use Adam (Kingma & Ba, 2014) as an optimizer with a starting learning rate of 0.0002, which we reduce in epoch 200, 250, and 300 by a factor of 0.25. ... An epoch consists of 1000 batches a 128 samples. We use early stopping to select the best parameters of the networks. ... We use the 34-layer Res Net architecture (He et al., 2015) but start with 16 feature maps instead of 64. The DCNNs are evaluated on the mean Hamming distance (MHD) i.e. the expected value of bits decoded wrong. ... For us γ = 15 worked well. |