LaRS Evaluator

Examples of scenes in the LaRS benchmark.

This is the evaluator code for the paper "LaRS: A Diverse Panoptic Maritime Obstacle Detection Dataset and Benchmark" presented at ICCV 2023. It can be used to evaluate semantic segmentation and panoptic segmentation predictions with the LaRS ground-truth annotations.

Currently only the GT of the training and validation sets is publicly available. For evaluation on the LaRS test set, please submit your submissions through our evaluation server.

Setup

1. Install requirements into your python environment

   pip install -r requirements.txt

2. For each of the evaluation tracks (semantic segmentation, panoptic segmentation) the evaluator expects a prediction root dir, where predictions will be placed.
   Configure paths to the dataset and predictions root in config files for your version of LaRS (e.g. lars_val_semantic.yaml).

Usage

1. Place the predictions of your methods into <prediction_root_dir>/<method_name> The method dir contains PNG files with predictions for all test images:
   • Semantic segmentation: The PNG file contains the predicted segmentation in RGB format, following the color coding of classes specified in the configuration file (e.g. lars_val_semantic.yaml). By default this is:
     • sky: [90, 75, 164]
     • water: [41, 167, 224]
     • obstacle: [247, 195, 37]
     • Alternatively you may use the lars_val_semantic_lbl.yaml config to evaluate predictions encoded as class ids (0 = obstacles, 1 = water, 2 = sky). Note, however, that the online evaluator expects predictions in the color-coded format.
   • Panoptic segmentation: The PNG file contains RGB coded class and instance predictions. The format follows LaRS GT masks: class id is stored in the R component, while instance ids are stored in the G and B components.
2. Run evaluation:

   $ python evaluate.py path/to/config.yaml <method_name>

Note

Result files with various statistics will be placed in the configured directory (results/v1.0.0/<track>/<method> by default).

Evaluation server

You can evaluate your methods on the LaRS test set through our online evaluation server. You need to create an account to submit your results.

The same prediction formats for semantic and panoptic segmentation tasks are expected by the server. The predictions of a single method need to be stored into a .zip archive and submitted to the server. More information >

Starter packs

We also provide the following starter packs for training and inferring on the LaRS dataset.

Semantic segmentation

Result files

Semantic segmentation

Results for semantic segmentation methods inlcude the following files:

• summary.csv: Overall results (IoU, water-edge accuracy, detection F1)
• frames.csv: Per frame metrics (number of TP, FP and FN, IoU, ...)
• segments.csv: Segment-wise results (TP coverage, FP area, FN area, ...)

Panoptic segmentation

Results for semantic segmentation methods inlcude the following files:

• summary.csv: Overall results (PQ, RQ, SQ, semantic metrics)
• frames.csv: Per frame metrics
• segments.csv: Segment-wise results (TPs, FPs, FNs, areas, bboxes)
• segments_agnostic.csv: Segment-wise results for obstacle-class-agnostic case
• segments_sem.csv: Segment-wise results from semantic segmentation evaluation
• obst_csl.csv: Matched segments (GT and pred) categories and IoU -> for confusion matrix
• obst_cls_agnostic.csv: Matched segments categories and IoU for obstacle-class-agnostic case

Citation

If you use LaRS, please cite our paper.

@InProceedings{Zust2023LaRS,
  title={LaRS: A Diverse Panoptic Maritime Obstacle Detection Dataset and Benchmark},
  author={{\v{Z}}ust, Lojze and Per{\v{s}}, Janez and Kristan, Matej},
  booktitle={International Conference on Computer Vision (ICCV)},
  year={2023}
}