For details about the model architecture, dataset, training procedure, and evaluation protocol, please refer to the original project and associated paper.
Available Checkpoints
The repository currently provides the following model configurations:
Directory
Description
CE_Adam
Cross-Entropy loss + Adam optimizer
CE_AdamW
Cross-Entropy loss + AdamW optimizer
CE_AdamW_sc
Cross-Entropy loss + AdamW optimizer + scheduler
FL_AdamW
Focal Loss + AdamW optimizer
FL_AdamW_sc
Focal Loss + AdamW optimizer + scheduler
LDAM
Label-Distribution-Aware Margin loss
LDAM_sc
Label-Distribution-Aware Margin loss + scheduler
WCE_AdamW
Weighted Cross-Entropy loss + AdamW optimizer
WCE_AdamW_sc
Weighted Cross-Entropy loss + AdamW optimizer + scheduler
Naming Convention
Model directories follow the naming convention:
<loss>_<optimizer>[_<scheduler>]
The scheduler component is optional.
For example:
CE_Adam
indicates a model trained with Cross-Entropy loss and the Adam optimizer without a learning-rate scheduler.
CE_AdamW_sc
indicates a model trained with Cross-Entropy loss and the AdamW optimizer with a learning-rate scheduler.
The suffix sc indicates that a scheduler was used during training.
Baseline Model
The CE_Adam configuration is the baseline model.
It uses:
Loss: Cross-Entropy (CE)
Optimizer: Adam
Scheduler: None
The other checkpoints represent alternative training configurations that can be compared against this baseline.
Checkpoint Files
Depending on the training configuration, each directory may contain one or more PyTorch checkpoint files, such as:
The checkpoint corresponding to the best-performing model observed during training according to the training/evaluation procedure.
final_model
The checkpoint saved at the end of training.
*_ddp.pth
These checkpoints were produced using Distributed Data Parallel (DDP) training.
The exact loading procedure may therefore depend on how the model was trained and implemented in the original project.
Reproducibility
This repository provides pretrained model weights corresponding to the training configurations used in the NACTI Species Recognition project.
For complete reproducibility, including model architecture, dataset preparation, training code, and evaluation procedures, please refer to the original project repository and associated publication.
The model weights in this repository are released under the MIT License.
Please also refer to the license and terms associated with the original project and its training data before redistributing or using the models in downstream applications.
Citation
If you use these pretrained models in your research, please cite the original NACTI Species Recognition work:
bibtex
1@conference{visapp26,
2author={Zehua Liu and Tilo Burghardt},
3title={Long-Tailed Species Recognition in the NACTI Wildlife Dataset},
4booktitle={Proceedings of the 21st International Conference on Computer Vision Theory and Applications - Volume 2: VISAPP},
5year={2026},
6pages={357-365},
7publisher={SciTePress},
8organization={INSTICC},
9doi={10.5220/0014342200004084},
10isbn={978-989-758-804-4},
11}
Please replace the citation above with the official bibliographic information from the original publication.
