edgefirst_client/coco/

reader.rs

// SPDX-License-Identifier: Apache-2.0
// Copyright © 2025 Au-Zone Technologies. All Rights Reserved.

//! Streaming COCO JSON/ZIP readers.
//!
//! Provides memory-efficient reading of COCO annotation files from JSON files
//! or ZIP archives without requiring full extraction.

use super::types::CocoDataset;
use crate::Error;
use std::{
    collections::HashSet,
    fs::File,
    io::{BufReader, Read},
    path::Path,
};

/// Options for COCO reading.
#[derive(Debug, Clone, Default)]
pub struct CocoReadOptions {
    /// If true, validate all annotations during reading.
    pub validate: bool,
    /// Maximum number of images to read (0 = unlimited).
    pub max_images: usize,
    /// Filter by category names (empty = all).
    pub category_filter: Vec<String>,
}

/// Streaming COCO reader for large datasets.
///
/// Supports reading from JSON files and ZIP archives.
///
/// # Example
///
/// ```rust,no_run
/// use edgefirst_client::coco::CocoReader;
///
/// let reader = CocoReader::new();
/// let dataset = reader.read_json("annotations/instances_val2017.json")?;
/// println!("Loaded {} images", dataset.images.len());
/// # Ok::<(), edgefirst_client::Error>(())
/// ```
pub struct CocoReader {
    options: CocoReadOptions,
}

impl CocoReader {
    /// Create a new COCO reader with default options.
    pub fn new() -> Self {
        Self {
            options: CocoReadOptions::default(),
        }
    }

    /// Create a new COCO reader with custom options.
    pub fn with_options(options: CocoReadOptions) -> Self {
        Self { options }
    }

    /// Read COCO dataset from a JSON file.
    ///
    /// # Arguments
    /// * `path` - Path to the COCO JSON annotation file
    ///
    /// # Returns
    /// Parsed `CocoDataset` structure
    pub fn read_json<P: AsRef<Path>>(&self, path: P) -> Result<CocoDataset, Error> {
        let file = File::open(path.as_ref())?;
        let reader = BufReader::with_capacity(64 * 1024, file);
        let mut dataset: CocoDataset = serde_json::from_reader(reader)?;
        fill_missing_file_names(&mut dataset);

        if self.options.validate {
            validate_dataset(&dataset)?;
        }

        Ok(self.apply_filters(dataset))
    }

    /// Read COCO annotations from a ZIP file.
    ///
    /// Looks for annotation JSON files in standard COCO locations:
    /// - `annotations/instances_*.json`
    /// - `annotations/*.json`
    /// - Root level `*.json` files
    ///
    /// # Arguments
    /// * `path` - Path to the ZIP archive containing annotations
    ///
    /// # Returns
    /// Merged `CocoDataset` from all annotation files found
    pub fn read_annotations_zip<P: AsRef<Path>>(&self, path: P) -> Result<CocoDataset, Error> {
        let file = File::open(path.as_ref())?;
        let mut archive = zip::ZipArchive::new(file)?;

        let mut merged = CocoDataset::default();

        for i in 0..archive.len() {
            let mut entry = archive.by_index(i)?;
            let name = entry.name().to_string();

            // Only process JSON files containing annotations
            if name.ends_with(".json") && name.contains("instances") {
                let mut contents = String::new();
                entry.read_to_string(&mut contents)?;

                let mut dataset: CocoDataset = serde_json::from_str(&contents)?;
                fill_missing_file_names(&mut dataset);
                merge_datasets(&mut merged, dataset);
            }
        }

        if self.options.validate {
            validate_dataset(&merged)?;
        }

        Ok(self.apply_filters(merged))
    }

    /// List image files in a COCO ZIP or folder.
    ///
    /// # Arguments
    /// * `path` - Path to COCO images folder or ZIP archive
    ///
    /// # Returns
    /// Vector of `(relative_path, absolute_path)` for each image
    pub fn list_images<P: AsRef<Path>>(
        &self,
        path: P,
    ) -> Result<Vec<(String, std::path::PathBuf)>, Error> {
        let path = path.as_ref();
        let mut images = Vec::new();

        if path.is_dir() {
            // Walk directory
            for entry in walkdir::WalkDir::new(path)
                .into_iter()
                .filter_map(|e| e.ok())
                .filter(|e| e.file_type().is_file())
            {
                let filename = entry.file_name().to_string_lossy().to_lowercase();
                if filename.ends_with(".jpg")
                    || filename.ends_with(".jpeg")
                    || filename.ends_with(".png")
                {
                    let rel_path = entry
                        .path()
                        .strip_prefix(path)
                        .unwrap_or(entry.path())
                        .to_string_lossy()
                        .to_string();
                    images.push((rel_path, entry.path().to_path_buf()));
                }
            }
        } else if path.extension().is_some_and(|e| e == "zip") {
            // List from ZIP
            let file = File::open(path)?;
            let mut archive = zip::ZipArchive::new(file)?;

            for i in 0..archive.len() {
                let entry = archive.by_index(i)?;
                let name = entry.name().to_string();
                let name_lower = name.to_lowercase();

                if !entry.is_dir()
                    && (name_lower.ends_with(".jpg")
                        || name_lower.ends_with(".jpeg")
                        || name_lower.ends_with(".png"))
                {
                    images.push((name.clone(), path.join(&name)));
                }
            }
        }

        Ok(images)
    }

    /// Read a single image from a ZIP archive.
    ///
    /// # Arguments
    /// * `zip_path` - Path to the ZIP archive
    /// * `image_name` - Name of the image file within the archive
    ///
    /// # Returns
    /// Raw image bytes
    pub fn read_image_from_zip<P: AsRef<Path>>(
        &self,
        zip_path: P,
        image_name: &str,
    ) -> Result<Vec<u8>, Error> {
        let file = File::open(zip_path.as_ref())?;
        let mut archive = zip::ZipArchive::new(file)?;

        let mut entry = archive.by_name(image_name)?;
        let mut buffer = Vec::with_capacity(entry.size() as usize);
        entry.read_to_end(&mut buffer)?;

        Ok(buffer)
    }

    /// Apply filters from options to the dataset.
    fn apply_filters(&self, mut dataset: CocoDataset) -> CocoDataset {
        // Apply max_images filter
        if self.options.max_images > 0 && dataset.images.len() > self.options.max_images {
            let image_ids: HashSet<_> = dataset
                .images
                .iter()
                .take(self.options.max_images)
                .map(|i| i.id)
                .collect();

            dataset.images.truncate(self.options.max_images);
            dataset
                .annotations
                .retain(|a| image_ids.contains(&a.image_id));
        }

        // Apply category filter
        if !self.options.category_filter.is_empty() {
            let category_ids: HashSet<_> = dataset
                .categories
                .iter()
                .filter(|c| self.options.category_filter.contains(&c.name))
                .map(|c| c.id)
                .collect();

            dataset
                .categories
                .retain(|c| self.options.category_filter.contains(&c.name));
            dataset
                .annotations
                .retain(|a| category_ids.contains(&a.category_id));
        }

        dataset
    }
}

impl Default for CocoReader {
    fn default() -> Self {
        Self::new()
    }
}

/// Validate a COCO dataset for consistency.
fn validate_dataset(dataset: &CocoDataset) -> Result<(), Error> {
    let image_ids: HashSet<_> = dataset.images.iter().map(|i| i.id).collect();
    let category_ids: HashSet<_> = dataset.categories.iter().map(|c| c.id).collect();

    for ann in &dataset.annotations {
        if !image_ids.contains(&ann.image_id) {
            return Err(Error::CocoError(format!(
                "Annotation {} references non-existent image_id {}",
                ann.id, ann.image_id
            )));
        }

        if !category_ids.contains(&ann.category_id) {
            return Err(Error::CocoError(format!(
                "Annotation {} references non-existent category_id {}",
                ann.id, ann.category_id
            )));
        }

        // Validate bbox
        if ann.bbox[2] <= 0.0 || ann.bbox[3] <= 0.0 {
            return Err(Error::CocoError(format!(
                "Annotation {} has invalid bbox dimensions",
                ann.id
            )));
        }
    }

    Ok(())
}

/// Derive a COCO-style relative path from an LVIS `coco_url`.
///
/// LVIS reuses COCO 2017 images and only records the URL (for example
/// `http://images.cocodataset.org/val2017/000000397133.jpg`). Everything
/// after the host is the same relative path COCO stores in `file_name`,
/// so we strip the scheme and host and return the remainder.
///
/// Downstream code joins the returned value onto an images base directory
/// (e.g. `images_dir.join(file_name)`), so the result is sanitised to
/// guarantee it stays inside that base:
///
/// - Leading `/` (absolute path) is rejected — `Path::join` would replace
///   the base entirely on Unix.
/// - Any `..` segment is rejected — `Path::join` preserves parent-dir
///   traversals and would escape the base.
/// - Windows drive prefixes (`C:\...`) are rejected — `Path::join` would
///   replace the base on Windows.
///
/// Sanitisation failures return `None`; callers leave `file_name` empty,
/// and later pipeline stages surface a clearer error.
fn derive_file_name_from_coco_url(url: &str) -> Option<String> {
    let after_scheme = url.split_once("://").map(|(_, rest)| rest).unwrap_or(url);
    let (_host, path) = after_scheme.split_once('/')?;
    if path.is_empty() {
        return None;
    }

    // Reject absolute paths. A leading `/` would make `Path::join` discard
    // the images base directory and resolve relative to the filesystem root.
    if path.starts_with('/') {
        return None;
    }

    // Reject any `..` segment. `Path::join` does not normalise parent
    // components, so `images_dir.join("../etc/passwd")` escapes the base.
    // Also reject literal backslashes — `Path::components` treats `/` as
    // the only separator on Unix, so `..\\foo` would slip past a naive
    // component check on a Linux build.
    //
    // Reject any `:` — `Component::Prefix` only fires on Windows builds,
    // so `C:\\Windows` reads as a single Normal component on Linux and
    // would be joined onto the base, only to blow up once the resulting
    // path hits a Windows host. COCO filenames never contain colons, so
    // this is a safe, portable way to reject drive-letter prefixes.
    if path.contains('\\') || path.contains(':') {
        return None;
    }
    let as_path = Path::new(path);
    for component in as_path.components() {
        use std::path::Component;
        match component {
            Component::Normal(_) => continue,
            // Any of these indicate a non-relative path or a traversal.
            Component::RootDir
            | Component::Prefix(_)
            | Component::ParentDir
            | Component::CurDir => return None,
        }
    }

    Some(path.to_string())
}

/// Populate empty `file_name` fields from `coco_url` (LVIS compatibility).
///
/// LVIS annotation JSONs omit `file_name`, relying on consumers to parse
/// `coco_url`. This runs once immediately after deserialisation so every
/// downstream stage (index building, Arrow conversion, studio upload) can
/// continue to assume `file_name` is populated.
fn fill_missing_file_names(dataset: &mut CocoDataset) {
    for image in &mut dataset.images {
        if !image.file_name.is_empty() {
            continue;
        }
        if let Some(derived) = image
            .coco_url
            .as_deref()
            .and_then(derive_file_name_from_coco_url)
        {
            image.file_name = derived;
        }
    }
}

/// Infer group name from COCO annotation filename.
///
/// Extracts the split name from standard COCO naming conventions:
/// - `instances_train2017.json` → `"train"`
/// - `instances_val2017.json` → `"val"`
/// - `instances_test2017.json` → `"test"`
/// - `person_keypoints_train2017.json` → `"train"`
///
/// # Arguments
/// * `filename` - The annotation file name
///
/// # Returns
/// The inferred group name if extraction succeeds
pub fn infer_group_from_filename(filename: &str) -> Option<String> {
    let stem = Path::new(filename).file_stem()?.to_str()?;

    // Try common COCO patterns
    // Pattern: instances_<group><year>.json
    if let Some(rest) = stem.strip_prefix("instances_") {
        let group = rest.trim_end_matches(char::is_numeric);
        if !group.is_empty() {
            return Some(group.to_string());
        }
    }

    // Pattern: person_keypoints_<group><year>.json
    if let Some(rest) = stem.strip_prefix("person_keypoints_") {
        let group = rest.trim_end_matches(char::is_numeric);
        if !group.is_empty() {
            return Some(group.to_string());
        }
    }

    // Pattern: captions_<group><year>.json
    if let Some(rest) = stem.strip_prefix("captions_") {
        let group = rest.trim_end_matches(char::is_numeric);
        if !group.is_empty() {
            return Some(group.to_string());
        }
    }

    // Pattern: panoptic_<group><year>.json
    if let Some(rest) = stem.strip_prefix("panoptic_") {
        let group = rest.trim_end_matches(char::is_numeric);
        if !group.is_empty() {
            return Some(group.to_string());
        }
    }

    // Fallback: look for train/val/test anywhere in the filename
    let lower = filename.to_lowercase();
    if lower.contains("train") {
        return Some("train".to_string());
    }
    if lower.contains("val") {
        return Some("val".to_string());
    }
    if lower.contains("test") {
        return Some("test".to_string());
    }

    None
}

/// Infer the group name from an image folder path.
///
/// Extracts the group from folder names like "train2017", "val2017",
/// "test2017". Strips trailing year numbers to get the group name.
///
/// # Examples
/// - `train2017/000000001.jpg` → "train"
/// - `val2017/000000002.jpg` → "val"
/// - `test2017/000000003.jpg` → "test"
/// - `custom_split/image.jpg` → "custom_split"
///
/// # Arguments
/// * `image_path` - Relative path to the image (e.g.,
///   "train2017/000000001.jpg")
///
/// # Returns
/// Inferred group name, or None if no folder component is found.
pub fn infer_group_from_folder(image_path: &str) -> Option<String> {
    let path = Path::new(image_path);

    // Get the parent folder name (e.g., "train2017" from "train2017/image.jpg")
    let folder = path.parent()?.file_name()?.to_str()?;

    if folder.is_empty() {
        return None;
    }

    // Strip trailing year numbers (e.g., "train2017" → "train")
    let group = folder.trim_end_matches(char::is_numeric);

    if group.is_empty() {
        // Folder was all digits, use original
        Some(folder.to_string())
    } else {
        Some(group.to_string())
    }
}

/// Read all COCO annotation files from a directory.
///
/// Discovers and reads annotation files from standard COCO directory
/// structures:
///
/// ```text
/// coco_dir/
/// ├── annotations/
/// │   ├── instances_train2017.json
/// │   └── instances_val2017.json
/// └── ...
/// ```
///
/// # Arguments
///
/// * `path` - Path to the COCO directory
/// * `options` - Read options
///
/// # Returns
///
/// Vector of `(CocoDataset, inferred_group)` pairs
pub fn read_coco_directory<P: AsRef<Path>>(
    path: P,
    options: &CocoReadOptions,
) -> Result<Vec<(CocoDataset, String)>, Error> {
    let path = path.as_ref();
    let mut results = Vec::new();

    // Look for annotation files
    let annotations_dir = path.join("annotations");
    let search_dirs: Vec<&Path> = if annotations_dir.is_dir() {
        vec![annotations_dir.as_path(), path]
    } else {
        vec![path]
    };

    for search_dir in search_dirs {
        if !search_dir.is_dir() {
            continue;
        }

        for entry in std::fs::read_dir(search_dir)? {
            let entry = entry?;
            let file_path = entry.path();

            if !file_path.is_file() {
                continue;
            }

            let filename = file_path.file_name().and_then(|s| s.to_str()).unwrap_or("");

            // Only process instance annotation files
            if filename.ends_with(".json") && filename.contains("instances") {
                let group =
                    infer_group_from_filename(filename).unwrap_or_else(|| "default".to_string());

                let reader = CocoReader::with_options(options.clone());
                let dataset = reader.read_json(&file_path)?;

                results.push((dataset, group));
            }
        }
    }

    if results.is_empty() {
        return Err(Error::MissingAnnotations(format!(
            "No COCO annotation files found in {}",
            path.display()
        )));
    }

    Ok(results)
}

/// Merge a source dataset into a target dataset.
fn merge_datasets(target: &mut CocoDataset, source: CocoDataset) {
    // Take info if not set
    if target.info.description.is_none() {
        target.info = source.info;
    }

    // Merge images (deduplicate by id)
    let existing_ids: HashSet<_> = target.images.iter().map(|i| i.id).collect();
    for image in source.images {
        if !existing_ids.contains(&image.id) {
            target.images.push(image);
        }
    }

    // Merge categories (deduplicate by id)
    let existing_cats: HashSet<_> = target.categories.iter().map(|c| c.id).collect();
    for cat in source.categories {
        if !existing_cats.contains(&cat.id) {
            target.categories.push(cat);
        }
    }

    // Merge annotations (always append - IDs should be globally unique)
    target.annotations.extend(source.annotations);

    // Merge licenses
    let existing_licenses: HashSet<_> = target.licenses.iter().map(|l| l.id).collect();
    for lic in source.licenses {
        if !existing_licenses.contains(&lic.id) {
            target.licenses.push(lic);
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::coco::{CocoAnnotation, CocoCategory, CocoImage};

    #[test]
    fn test_reader_default() {
        let reader = CocoReader::new();
        assert!(!reader.options.validate);
        assert_eq!(reader.options.max_images, 0);
        assert!(reader.options.category_filter.is_empty());
    }

    #[test]
    fn test_reader_with_options() {
        let options = CocoReadOptions {
            validate: true,
            max_images: 100,
            category_filter: vec!["person".to_string()],
        };
        let reader = CocoReader::with_options(options.clone());
        assert!(reader.options.validate);
        assert_eq!(reader.options.max_images, 100);
    }

    #[test]
    fn test_derive_file_name_from_coco_url() {
        assert_eq!(
            derive_file_name_from_coco_url(
                "http://images.cocodataset.org/val2017/000000397133.jpg"
            ),
            Some("val2017/000000397133.jpg".to_string())
        );
        assert_eq!(
            derive_file_name_from_coco_url(
                "https://images.cocodataset.org/train2017/000000000009.jpg"
            ),
            Some("train2017/000000000009.jpg".to_string())
        );
        assert_eq!(derive_file_name_from_coco_url("host-only"), None);
        assert_eq!(derive_file_name_from_coco_url("http://host/"), None);
    }

    #[test]
    fn test_derive_file_name_from_coco_url_rejects_traversal() {
        // Parent-dir segments must not survive — `Path::join` would otherwise
        // let `images_dir.join("../etc/passwd")` escape the images base.
        assert_eq!(
            derive_file_name_from_coco_url("http://host/../etc/passwd"),
            None
        );
        assert_eq!(
            derive_file_name_from_coco_url("http://host/val2017/../../etc/passwd"),
            None
        );
        // A lone `.` segment is also rejected (no legitimate reason to keep
        // it, and refusing keeps the sanitiser easy to reason about).
        assert_eq!(derive_file_name_from_coco_url("http://host/./foo.jpg"), None);
    }

    #[test]
    fn test_derive_file_name_from_coco_url_rejects_absolute_and_windows() {
        // Leading `/` after the host means the URL encoded an absolute
        // filesystem path (`http://host//etc/passwd`). `Path::join` drops
        // the base on Unix when joining an absolute path.
        assert_eq!(
            derive_file_name_from_coco_url("http://host//etc/passwd"),
            None
        );
        // Backslashes would slip past a naive Component check on Linux
        // because `/` is the only separator there.
        assert_eq!(
            derive_file_name_from_coco_url("http://host/val2017\\..\\..\\etc"),
            None
        );
        // Windows drive prefix — `Path::join` drops the base on Windows
        // when joining a path with a drive.
        assert_eq!(
            derive_file_name_from_coco_url("http://host/C:/Windows/System32"),
            None
        );
    }

    #[test]
    fn test_fill_missing_file_names_from_lvis_json() {
        // LVIS-style image record: no `file_name`, only `coco_url`.
        let json = r#"{
            "images": [
                {
                    "id": 397133,
                    "width": 640,
                    "height": 427,
                    "coco_url": "http://images.cocodataset.org/val2017/000000397133.jpg",
                    "neg_category_ids": [279, 899],
                    "not_exhaustive_category_ids": [914]
                }
            ],
            "annotations": [],
            "categories": []
        }"#;
        let mut dataset: CocoDataset = serde_json::from_str(json).unwrap();
        assert_eq!(dataset.images[0].file_name, "");
        fill_missing_file_names(&mut dataset);
        assert_eq!(dataset.images[0].file_name, "val2017/000000397133.jpg");
        // LVIS extension fields survive deserialisation.
        assert_eq!(
            dataset.images[0].neg_category_ids.as_deref(),
            Some(&[279u32, 899][..])
        );
    }

    #[test]
    fn test_fill_missing_file_names_preserves_existing() {
        // COCO-style records already carry `file_name`; we must not clobber them.
        let mut dataset = CocoDataset {
            images: vec![CocoImage {
                id: 1,
                width: 640,
                height: 480,
                file_name: "custom/path.jpg".to_string(),
                coco_url: Some("http://images.cocodataset.org/val2017/foo.jpg".to_string()),
                ..Default::default()
            }],
            ..Default::default()
        };
        fill_missing_file_names(&mut dataset);
        assert_eq!(dataset.images[0].file_name, "custom/path.jpg");
    }

    #[test]
    fn test_validate_dataset_valid() {
        let dataset = CocoDataset {
            images: vec![CocoImage {
                id: 1,
                width: 640,
                height: 480,
                file_name: "test.jpg".to_string(),
                ..Default::default()
            }],
            categories: vec![CocoCategory {
                id: 1,
                name: "person".to_string(),
                supercategory: None,
                ..Default::default()
            }],
            annotations: vec![CocoAnnotation {
                id: 1,
                image_id: 1,
                category_id: 1,
                bbox: [10.0, 20.0, 100.0, 80.0],
                area: 8000.0,
                iscrowd: 0,
                segmentation: None,
                score: None,
            }],
            ..Default::default()
        };

        assert!(validate_dataset(&dataset).is_ok());
    }

    #[test]
    fn test_validate_dataset_missing_image() {
        let dataset = CocoDataset {
            images: vec![],
            categories: vec![CocoCategory {
                id: 1,
                name: "person".to_string(),
                supercategory: None,
                ..Default::default()
            }],
            annotations: vec![CocoAnnotation {
                id: 1,
                image_id: 999, // Non-existent
                category_id: 1,
                bbox: [10.0, 20.0, 100.0, 80.0],
                ..Default::default()
            }],
            ..Default::default()
        };

        assert!(validate_dataset(&dataset).is_err());
    }

    #[test]
    fn test_merge_datasets() {
        let mut target = CocoDataset {
            images: vec![CocoImage {
                id: 1,
                width: 640,
                height: 480,
                file_name: "img1.jpg".to_string(),
                ..Default::default()
            }],
            categories: vec![CocoCategory {
                id: 1,
                name: "person".to_string(),
                supercategory: None,
                ..Default::default()
            }],
            annotations: vec![],
            ..Default::default()
        };

        let source = CocoDataset {
            images: vec![
                CocoImage {
                    id: 1, // Duplicate - should not be added
                    width: 640,
                    height: 480,
                    file_name: "img1.jpg".to_string(),
                    ..Default::default()
                },
                CocoImage {
                    id: 2, // New - should be added
                    width: 800,
                    height: 600,
                    file_name: "img2.jpg".to_string(),
                    ..Default::default()
                },
            ],
            categories: vec![CocoCategory {
                id: 2,
                name: "car".to_string(),
                supercategory: None,
                ..Default::default()
            }],
            annotations: vec![],
            ..Default::default()
        };

        merge_datasets(&mut target, source);

        assert_eq!(target.images.len(), 2);
        assert_eq!(target.categories.len(), 2);
    }

    #[test]
    fn test_apply_max_images_filter() {
        let reader = CocoReader::with_options(CocoReadOptions {
            max_images: 2,
            ..Default::default()
        });

        let dataset = CocoDataset {
            images: vec![
                CocoImage {
                    id: 1,
                    ..Default::default()
                },
                CocoImage {
                    id: 2,
                    ..Default::default()
                },
                CocoImage {
                    id: 3,
                    ..Default::default()
                },
            ],
            annotations: vec![
                CocoAnnotation {
                    id: 1,
                    image_id: 1,
                    ..Default::default()
                },
                CocoAnnotation {
                    id: 2,
                    image_id: 2,
                    ..Default::default()
                },
                CocoAnnotation {
                    id: 3,
                    image_id: 3,
                    ..Default::default()
                },
            ],
            ..Default::default()
        };

        let filtered = reader.apply_filters(dataset);
        assert_eq!(filtered.images.len(), 2);
        assert_eq!(filtered.annotations.len(), 2);
    }

    #[test]
    fn test_infer_group_from_filename_instances() {
        assert_eq!(
            infer_group_from_filename("instances_train2017.json"),
            Some("train".to_string())
        );
        assert_eq!(
            infer_group_from_filename("instances_val2017.json"),
            Some("val".to_string())
        );
        assert_eq!(
            infer_group_from_filename("instances_test2017.json"),
            Some("test".to_string())
        );
    }

    #[test]
    fn test_infer_group_from_filename_keypoints() {
        assert_eq!(
            infer_group_from_filename("person_keypoints_train2017.json"),
            Some("train".to_string())
        );
        assert_eq!(
            infer_group_from_filename("person_keypoints_val2017.json"),
            Some("val".to_string())
        );
    }

    #[test]
    fn test_infer_group_from_filename_captions() {
        assert_eq!(
            infer_group_from_filename("captions_train2017.json"),
            Some("train".to_string())
        );
        assert_eq!(
            infer_group_from_filename("captions_val2017.json"),
            Some("val".to_string())
        );
    }

    #[test]
    fn test_infer_group_from_filename_panoptic() {
        assert_eq!(
            infer_group_from_filename("panoptic_train2017.json"),
            Some("train".to_string())
        );
        assert_eq!(
            infer_group_from_filename("panoptic_val2017.json"),
            Some("val".to_string())
        );
    }

    #[test]
    fn test_infer_group_from_filename_fallback() {
        // Falls back to looking for train/val/test in filename
        assert_eq!(
            infer_group_from_filename("my_custom_train_annotations.json"),
            Some("train".to_string())
        );
        assert_eq!(
            infer_group_from_filename("validation_data.json"),
            Some("val".to_string())
        );
    }

    #[test]
    fn test_infer_group_from_filename_no_match() {
        // No recognizable pattern
        assert_eq!(infer_group_from_filename("annotations.json"), None);
        assert_eq!(infer_group_from_filename("data.json"), None);
    }
}