Note: This codebase has only been tested on Debian-based Linux systems. CUDA is required to install the dependencies and run experiments.

The Open-Insect Dataset

Open-Insect

The Open-Insect dataset with GBIF images is publicly avaiable at Open-Insect on hugggingface.

To download the images and generate metadata for training:

• Change download_dir to the directory where you want the downloaded dataset to be saved.
• The resize_size is the smaller edge of the image after resizing. Change resize_size accordingly. The default value is 224.
• If you do not want to resize the images, simply delete --resize_size 224 from the command. Without resizing, the downloaded images will require approximately 6TB of storage.
• Run

  bash download.sh
  

Once downloading finishes,

• Images will be saved under <download_dir>/images.
• Metadata for training and evaluation of each region will be saved as

  <download_dir>/metadata/<region>
  │   ├── test_id.csv
  │   ├── test_ood_local.csv
  │   ├── test_ood_non-local.csv
  │   ├── test_ood_non-moth.csv
  │   ├── train_aux.csv
  │   ├── train_id.csv
  │   ├── val_id.csv
  │   └── val_ood.csv
  

• Change data_dir, imglist_pth, and pre_size in the configs under configs/datasets accordingly before training or evaluation.

Open-Insect-BCI

The C-America O-BCI dataset is hosted separately at Open-Insect-BCI on huggingface.

Run

python download_bci.py --download_dir .

to download the BCI dataset to the current directory, or change the download_dir accordingly.

• Images will be saved under <download_dir>/images/bci.
• Metadata will be saved as <dowload_dir>/metadata/c-america/test_ood_bci.txt.

Requirements

Run the following commands to install dependencies.

conda create -n oi_env python=3.10

conda activate oi_env

pip install -e .

pip install libmr

The default batch size is 512 and the number of works is 16. With this setting, models can be trained with 1 RTX800 GPU with 48 GB memory, 16 CPUs (and 16 workers), and 100 GB CPU memory in total.

Arguments

Argument	Description	Possible Values / Examples
REGION	Specifies the geographical region of the dataset to evaluate.	ne-america (Northeastern America), w-europe (Western Europe), c-america (Central America)
METHOD	The training method.	See the list METHOD Options below.
POSTHOC_METHOD	The post-hoc open-set detection method applied to the trained classifier.	See the list POSTHOC_METHOD Options below.
NETWORK	The backbone network used in the model.	See the list NETWORK Options below.
CHECKPOINT_DIR	Path to the directory to save the trained model checkpoints.	Example: $HOME/weights

METHOD Options (See Table 2 in the paper for more details of the methods)

• basics - the basic classifier trained with Cross Entropy loss with only the closed-set
• conf_branch - ConfBranch
• logitnorm - LogitNorm
• godin - GODIN
• rotpred- RotPred
• oe - OE
• udg - UDG
• mixoe - MixOE
• energy - Energy
• extended - Extended
• novel_branch- NovelBranch

---

POSTHOC_METHOD Options (See Table 2 in the paper for more details of the post-hoc methods)

• Generic post-hoc methods
  • openmax — OpenMax
  • msp — MSP
  • temperature_scaling — TempScale
  • odin — ODIN
  • mds — MDS
  • mds_ensemble — MDSEns
  • rmds — RMDS
  • gram — Gram
  • ebo — EBO
  • gradnorm — GradNorm
  • react - ReAct
  • mls — MLS
  • klm — KLM
  • vim — VIM
  • knn — k-Nearest Neighbor in feature space
  • dice - DICE
  • rankfeat - RankFeat
  • ash — ASH
  • she — SHE
  • neco — NECO
  • fdbd — FDBD
  • rp_msp, rp_odin, rp_ebo, rp_gradnorm - RP_MSP, RP_ODIN, RP_EBO, RP_GradNorm
  • nci — NCI
• Post-hoc methods for a specific training method
  • conf_branch - To be used with METHOD: conf_branch
  • godin - To be used with METHOD: godin
  • rotpred - To be used with METHOD: rotpred

---

NETWORK Options

• conf_branch - for METHOD: conf_branch
• godin_net - for METHOD: godin
• rot_net - for METHOD: rotpred
• udg_net - for METHOD: udg
• extended_net - for METHOD: extended or novel_branch
• resnet50 - for all other methods

Training

Run the following command to train from scratch. The model checkpoint will be saved in ${CHECKPOINT_DIR}/${REGION}/${METHOD}/train_from_scratch/s${RANDOM_SEED}.

bash scripts/train.sh REGION METHOD NETWORK CHECKPOINT_DIR

Run the following command to fine-tune the CHECKPOINT. The fine-tuned model will be saved in CHECKPOINT_DIR.

bash scripts/finetune.sh REGION METHOD NETWORK CHECKPOINT_DIR CHECKPOINT

Evaluation

Evaluating ARPL

bash scripts/eval_arpl.sh REGION arpl msp arpl_net CHECKPOINT_DIR

Evaluating OpenGAN

bash scripts/eval_opengan.sh REGION opengan opengan resnet50 CHECKPOINT_DIR

Evaluating other methods

For all other methods, run

bash scripts/eval.sh REGION METHOD POSTHOC_METHOD NETWORK CHECKPOINT_DIR

For example, to evaluate the basic classifier for Central America with MSP using the checkpoint saved under $HOME/weights, run

bash scripts/eval.sh c-america basics msp resnet50 $HOME/weights

See scripts/test_eval_script.sh for more examples.

Possible errors

• OpenMax: This method requires predictions of the test set to cover all training species. Otherwise, the following error will occur: RuntimeError: torch.cat(): expected a non-empty list of Tensors.

Pre-trained checkpoints

Checkpoints can be downloaded from https://huggingface.co/yuyan-chen/open-insect-model-weights or by running

python download_pretrained_weights.py --weight_dir WEIGHT_DIR

Examples

Here are some minimal examples to test this codebase. First, activate the virtual environment by

conda activate oi_env

then download the pretrained weights by running

python download_pretrained_weights.py

The model should be saved under weights/c-america_resnet50_baseline.pth.

Training

To test training methods that do no require auxiliary data, run

bash scripts/examples/train.sh

To test training methods that require auxiliary data, run

bash scripts/examples/train_with_aux_data.sh

The training and validation accuracy are expected to be 0 after 2 epochs as there is only 1 image per speices in the training set, and the model is trained from scratch.

Evaluation

Run

bash scripts/examples/eval.sh

The outputs will be saved in results/open-insect-example/base/msp/ood.csv. You can compare the outputs with example/ood.csv.

Acknowledgement

This codebase is built using OpenOOD. We sincerely appreciate their efforts in making this valuable resource publicly available.

Citation

Our paper:

@inproceedings{
chen2025openinsect,
title={Open-Insect: Benchmarking Open-Set Recognition of Novel Species in Biodiversity Monitoring},
author={Yuyan Chen and Nico Lang and B. Christian Schmidt and Aditya Jain and Yves Basset and Sara Beery and Maxim Larriv{\'e}e and David Rolnick},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
year={2025},
url={https://openreview.net/forum?id=63Tia99ofI}
}