euler-eval

A comprehensive evaluation toolkit for comparing predicted depth maps, RGB images, and camera ray direction maps against ground truth, powered by euler_loading for flexible dataset loading.

Features

• Depth metrics: PSNR, SSIM, LPIPS, FID, KID, AbsRel, RMSE, Scale-Invariant Log Error, Normal Consistency, Depth Edge F1
• RGB metrics: PSNR, SSIM, LPIPS, FID, SCE (Structural Chromatic Error), Edge F1, Tail Errors (p95/p99), High-Frequency Energy Ratio, Depth-Binned Photometric Error
• Rays metrics: ρ_A (AUC of angular accuracy curve), Angular Error statistics and threshold percentages
• Points-3D metrics: for models that predict a per-pixel 3D point map and their own camera: Euclidean 3D error (EPE/RMSE/3D δ-accuracy), radial-vs-lateral error decomposition (depth vs camera-model attribution) with ρ_A, true-3D surface normals, 3D edge F1, and correspondence-free Chamfer / F-score. Native + similarity-gauge-aligned (Umeyama) metric spaces.
• Benchmark binning: Optional depth-range benchmark that subdivides metrics into square-root-scaled near/mid/far bins
• Sanity checking: Automatic validation of metric results against configurable thresholds, with detailed warning reports
• Sky masking: Optional exclusion of sky regions from metrics using GT segmentation
• Flexible dataset loading: Automatic loader resolution via euler_loading and ds-crawler index metadata
• Per-file and aggregate results: Outputs both per-image metrics and dataset-level aggregates to JSON, saved per-modality
• euler_train integration: Optional experiment logging via euler_train

Installation

Requires Python 3.9+.

uv pip install "euler-eval @ git+https://github.com/d-rothen/euler-parser.git"

# with euler_train logging support
uv pip install "euler-eval[logging] @ git+https://github.com/d-rothen/euler-parser.git"

# with clean-fid RGB FID backend support
uv pip install "euler-eval[fid] @ git+https://github.com/d-rothen/euler-parser.git"

Or install in editable mode:

pip install -e .

Dependencies

Core:

• numpy, scipy, Pillow
• torch, torchvision
• lpips
• tqdm
• euler-loading, ds-crawler

Optional:

• euler-train (install via [logging] extra)

Usage

The package provides a depth-eval console script:

depth-eval <config> [options]

It also provides a cache warmup helper for offline environments:

euler-eval.init

Or run directly:

python main.py <config> [options]

Before running on offline compute nodes, you can warm caches on a machine with network access:

HF_HOME=/shared/cache/hf \
TORCH_HOME=/shared/cache/hf/torch \
CLEANFID_CACHE_DIR=/shared/cache/clean-fid \
euler-eval.init

This pre-downloads:

• torchvision AlexNet weights
• torchvision Inception v3 weights
• LPIPS AlexNet weights
• the clean-fid inception checkpoint, if clean-fid is installed

Positional arguments

Options

Examples

# Evaluate with default settings (auto-selects CUDA when available)
depth-eval config.json --batch-size 32

# Evaluate with sky masking enabled (requires gt.segmentation or gt.semantic_segmentation)
depth-eval config.json --mask-sky -v

# Skip RGB evaluation, only evaluate depth
depth-eval config.json --skip-rgb

# Disable sanity checking
depth-eval config.json --no-sanity-check

# Disable depth alignment
depth-eval config.json --depth-alignment none

# Force affine scale+shift alignment on all depth predictions
depth-eval config.json --depth-alignment affine

# Use clean-fid for RGB FID computation
depth-eval config.json --rgb-fid-backend clean-fid

# Benchmark depth and RGB metrics within a depth range (near/mid/far bins)
depth-eval config.json --benchmark-depth-range 0.01 80.0

# Evaluate dense depth predictions against sparse pointcloud GT
depth-eval example_sparse_depth_config.json --skip-rgb --skip-rays

# Skip rays evaluation
depth-eval config.json --skip-rays

# Evaluate per-pixel 3D point maps (predict-your-own-camera models)
depth-eval example_points_3d_config.json --skip-depth --skip-rgb --skip-rays

# Align relative point maps to GT with a similarity (Umeyama) gauge
depth-eval example_points_3d_config.json --points-3d-alignment similarity

Benchmark Depth Bins

--benchmark-depth-range MIN MAX first filters valid GT depth pixels to MIN <= depth <= MAX. The near, mid, and far bins are then computed as three equal-width intervals in square-root depth space:

sqrt_min = sqrt(MIN)
sqrt_max = sqrt(MAX)
step = (sqrt_max - sqrt_min) / 3
near_max = (sqrt_min + step)^2
mid_max = (sqrt_min + 2 * step)^2

The interval bounds are near=[MIN, near_max), mid=[near_max, mid_max), and far=[mid_max, MAX]. For --benchmark-depth-range 0.01 80.0, this gives:

Configuration

config.json

Defines GT modalities, prediction datasets to evaluate, and optional euler_train logging. See example_config.json. For sparse pointcloud depth GT, see example_sparse_depth_config.json.

{
  "euler_train": {
    "dir": "runs/my_project"
  },
  "gt": {
    "rgb":          { "path": "/data/gt/rgb" },
    "depth":        { "path": "/data/gt/depth" },
    "rays":         { "path": "/data/gt/rays" },
    "segmentation": { "path": "/data/gt/segmentation" },
    "calibration":  { "path": "/data/gt/calibration" }
  },
  "datasets": [
    {
      "name": "model_a",
      "rgb":   { "path": "/data/model_a/rgb" },
      "depth": { "path": "/data/model_a/depth" },
      "rays":  { "path": "/data/model_a/rays" },
      "output_file": "/path/to/output/model_a_eval.json"
    },
    {
      "name": "model_b_depth_only",
      "depth": { "path": "/data/model_b/depth" }
    },
    {
      "name": "model_c_rgb_only",
      "rgb": { "path": "/data/model_c/rgb" }
    }
  ]
}

Each modality entry can optionally include a split field to select a specific split from the dataset (e.g. { "path": "/data/gt/depth", "split": "test" }). euler-loading inline selectors are also accepted in path, such as /data/muses.zip:test or /data/muses.zip:test#scope=rgb.

For sparse pointcloud depth evaluation, use gt.sparse_depth instead of gt.depth. The prediction uses a dense depth-like map under datasets[].depth, datasets[].relative_depth, or datasets[].affine_depth. The evaluator projects the sparse GT point cloud into the prediction image plane using gt.intrinsics and gt.camera_extrinsics, then computes pointwise depth metrics only at projected valid pixels. For MUSES through euler_loading.loaders.muses, gt.camera_extrinsics normally resolves to the direct lidar2rgb transform and no extra lidar pose is needed. If a dataset exposes separate lidar and camera poses in a shared frame instead, provide optional gt.lidar_extrinsics; the evaluator composes inv(camera_pose) @ lidar_pose before projection.

Sparse depth does not require segmentation GT. gt.segmentation and gt.semantic_segmentation are optional aliases for the same sky-mask source. They are loaded only when --mask-sky is set; in sparse depth evaluation that mask excludes sky pixels from projected-point metrics and from scale/shift fitting. Without --mask-sky, the segmentation entry is not loaded or used.

GT section

* At least one of gt.rgb.path, gt.depth.path, gt.sparse_depth.path, gt.rays.path, or gt.points_3d.path is required.

Prediction datasets

Each entry in datasets can include rgb, one dense depth-like prediction (depth, relative_depth, or affine_depth), rays, or any combination:

* At least one of rgb.path, depth.path, relative_depth.path, affine_depth.path, rays.path, or points_3d.path is required. Use only one dense depth-like entry per prediction dataset.

When evaluating a points_3d prediction, the GT point map comes from one of two sources:

1. Explicit GT point map — gt.points_3d.path (a stored (H,W,3) map).
2. Synthesized from depth (fallback) — if gt.points_3d is absent but gt.depth.path and intrinsics (gt.intrinsics.path or gt.calibration.path) are present, the evaluator unprojects GT depth with the intrinsics into a GT point map on the fly. This lets you score a points_3d prediction against a conventional depth + intrinsics ground truth without precomputing point maps. Whether the GT depth is radial or planar is read from its dataset metadata (radial_depth).

GT intrinsics under gt.calibration (or gt.intrinsics) are also used to auto-detect the FoV domain for the angular ρ_A sub-metric; for the explicit-point-map path they are optional and default to the lfov threshold when absent, but for the depth-synthesis fallback they are required. The gauge alignment is controlled by --points-3d-alignment (see Points-3D gauge alignment).

When relative_depth or affine_depth is used, --depth-alignment auto_affine treats the entry as declared non-metric depth and runs scale/shift alignment even if the raw value range is not normalized. --depth-alignment none still disables calibration.

euler_train section (optional)

When present, evaluation results are logged to an euler_train run. Requires the euler-train package to be installed (pip install euler-eval[logging]).

euler_train auto-detects whether the path is a run directory by checking for meta.json. When resuming an existing run, the run is detached after evaluation (the run remains active for further use). When a new run is created, it is finished upon completion.

Loader resolution

Loaders are resolved automatically by euler_loading from each dataset directory's ds-crawler index metadata. The index's euler_loading.loader and euler_loading.function fields determine which loader module and function to use (e.g. "vkitti2" maps to euler_loading.loaders.gpu.vkitti2).

No manual loader selection is required. Each dataset directory declares its own loader through its ds-crawler configuration.

Dataset metadata (e.g. radial_depth, rgb_range, sparse point columns, and coordinate units) is read automatically from each dataset's ds-crawler metadata via get_modality_metadata(). Depth and point-cloud coordinates are assumed to already be in meters.

Dataset Metadata

Each dataset directory or archive must contain ds-crawler metadata artifacts, typically generated under .ds_crawler/:

dataset-root/
  .ds_crawler/
    dataset-head.json
    ds-crawler.json
    index.json

When one physical root or archive contains several logical modalities, the artifacts may instead be scoped under .ds_crawler/<modality>/ with a .ds_crawler/scopes.json manifest. euler-eval passes modality scopes to euler-loading so paths such as /data/muses.zip:test can load .ds_crawler/rgb/index.json, .ds_crawler/depth/index.json, and related scoped artifacts from the same archive.

GT and prediction datasets are matched by hierarchy path and file ID through MultiModalDataset.

metrics_config.json

Controls sanity check thresholds. See metrics_config.json for all available options. When --metrics-config is not specified, the tool auto-detects metrics_config.json at the project root. If not found, built-in defaults are used.

Metrics

Depth metrics

Sparse Depth Metrics

Sparse pointcloud GT evaluation reports only metrics that remain meaningful at isolated projected points. Dense image-quality and geometric metrics such as SSIM, LPIPS, FID, normals, and edge F1 are intentionally skipped. Serialized eval.json sparse-depth metrics use the namespace root sparsedepth (without an underscore), so flattened metric names match metricSet.metricNamespace = "sparsedepth.eval".

RGB metrics

Rays metrics

Points-3D metrics

For per-pixel 3D point maps P(u,v) = (X,Y,Z) in camera-frame metres. Metrics are emitted in two gauge spaces (native raw prediction, metric after alignment) with a points_3d canonical alias, keyed as points3d.eval.{native,metric}.{category}…. Categories:

Points-3D gauge alignment

Point maps from different model families live in different frames, so points_3d resolves an unknown similarity gauge (the 3D generalization of depth's scale-and-shift) before the metric-space comparison:

• none — score the raw points as-is (metric, predict-your-own-camera models).
• scale — fit a single global scalar s (least squares).
• similarity — fit a 7-DoF Umeyama transform s·R·p + t over the known per-pixel correspondences (relative / free-frame models).
• auto (default) — apply similarity for declared-relative predictions, else none.

The angular error_decomposition ray metrics and rho_a are computed on the native point-map directions (the camera-faithful frame); the radial/lateral and Euclidean metrics are reported per space.

Output

Results are saved as JSON per modality per prediction dataset (one file for depth or sparse depth, one for RGB, one for rays). Default path: eval.json inside each modality's dataset path, unless overridden by output_file in the config.

For RGB FID, two backends are available:

• builtin: in-process Inception-based implementation in this repository.
• clean-fid: delegates folder-vs-folder FID computation to clean-fid. This backend requires installing the optional fid extra and is recommended when you need scores closer to standard published FID numbers.

When --rgb-fid-backend clean-fid is used, euler-eval will honor CLEANFID_CACHE_DIR if set:

• If CLEANFID_CACHE_DIR/inception-2015-12-05.pt exists, it is staged into the location clean-fid expects before evaluation.
• If it does not exist and the machine is online, euler-eval asks clean-fid to download it into CLEANFID_CACHE_DIR.
• Without CLEANFID_CACHE_DIR, clean-fid falls back to its own default local path handling.

Output structure

{
  "depth_native": { "...": "native model depth space, if diagnostically meaningful" },
  "depth_metric": { "...": "metric depth space, if available" },
  "depth": {
    "...": "canonical alias of depth_metric when present, else depth_native"
  },
  "sparsedepth": {
    "eval": { "...": "serialized sparse-depth metrics when gt.sparse_depth is used" }
  },
  "points3d": {
    "eval": {
      "native": { "...": "raw-prediction point-map metrics" },
      "metric": { "...": "gauge-aligned point-map metrics, when alignment ran" }
    }
  },
  "rgb": {
    "...": "..."
  },
  "per_file_metrics": {
    "children": {
      "scene_01": {
        "children": {
          "camera_0": {
            "files": [
              {
                "id": "frame_0001",
                "metrics": {
                  "depth": { "...": "canonical alias" },
                  "depth_native": { "...": "native, when emitted" },
                  "depth_metric": { "...": "metric, when emitted" },
                  "sparsedepth": { "eval": { "...": "serialized sparse-depth metrics" } },
                  "rgb": { "...": "..." }
                }
              }
            ]
          }
        }
      }
    }
  }
}

For depth outputs:

• depth_native: the model's native depth space after spatial/radial preprocessing, emitted only when it is diagnostically distinct.
• depth_metric: the comparable metric-depth branch. This is either the native prediction itself or the calibrated scale-shift result.
• depth: canonical alias of depth_metric when available, otherwise depth_native.
• standard: explicit monocular-depth metrics with three reducers: image_mean, image_median, and pixel_pool.

For sparse depth outputs, the internal Python result dict still uses sparse_depth_native, sparse_depth_metric, and sparse_depth, but serialized eval.json metric paths are rooted at sparsedepth.eval to satisfy external namespace validation. Only standard and depth_metrics categories are present.

For points_3d outputs, the internal Python result dict uses points_3d_native, points_3d_metric, and points_3d (canonical alias), but serialized eval.json metric paths are rooted at points3d.eval (no underscore) to satisfy the namespace first-segment rule, mirroring sparsedepth. The metric space is emitted only when a gauge alignment is applied; otherwise only native is present and aliased.

Previous single-depth structure (kept under depth) is:

{
  "depth": {
    "image_quality": {
      "psnr": 28.5,
      "ssim": 0.92,
      "lpips": 0.08,
      "fid": 12.3,
      "kid_mean": 0.005,
      "kid_std": 0.002
    },
    "standard": {
      "image_mean": {
        "absrel": 0.08,
        "sqrel": 0.04,
        "mae": 0.62,
        "rmse": 1.20,
        "rmse_log": 0.11,
        "log10": 0.04,
        "silog": 0.08,
        "delta1": 0.91,
        "delta2": 0.97,
        "delta3": 0.99
      },
      "image_median": {
        "absrel": 0.07,
        "sqrel": 0.03,
        "mae": 0.58,
        "rmse": 1.10,
        "rmse_log": 0.10,
        "log10": 0.04,
        "silog": 0.07,
        "delta1": 0.92,
        "delta2": 0.98,
        "delta3": 0.99
      },
      "pixel_pool": {
        "absrel": 0.08,
        "sqrel": 0.04,
        "mae": 0.61,
        "rmse": 1.18,
        "rmse_log": 0.11,
        "log10": 0.04,
        "silog": 0.08,
        "delta1": 0.91,
        "delta2": 0.97,
        "delta3": 0.99
      }
    },
    "depth_metrics": {
      "absrel": { "median": 0.05, "p90": 0.12 },
      "rmse":   { "median": 1.2, "p90": 3.1 },
      "silog":  { "mean": 0.08, "median": 0.06, "p90": 0.15 }
    },
    "geometric_metrics": {
      "normal_consistency": {
        "mean_angle": 12.3,
        "median_angle": 9.8,
        "percent_below_11_25": 55.2,
        "percent_below_22_5": 82.1,
        "percent_below_30": 91.5
      },
      "depth_edge_f1": {
        "precision": 0.72,
        "recall": 0.68,
        "f1": 0.70
      }
    },
    "dataset_info": {
      "num_pairs": 500,
      "gt_name": "GT",
      "pred_name": "model_a"
    }
  },
  "rgb": { "...": "unchanged" }
}

Sanity check report

When sanity checking is enabled (the default), a sanity_check_report.json is saved to the current working directory containing warnings grouped by metric type.

License

MIT