Unmarked Anything

Segment and Depth Anything 3 for Camera Trap Distance Sampling

This repository provides a workflow for camera trap–based distance sampling that integrates segmentation and monocular depth estimation to localise, track, and estimate animal-to-camera distances for a specified target species. The framework combines the wildlife-adapted SAM-3 (SA-FARI) with Depth Anything 3 to generate frame-level detections, segmentation masks, object tracks, and depth-derived distance signals from camera trap video.

Core Technologies

• SAM-3 (SA-FARI) for wildlife-specific segmentation/tracking: https://arxiv.org/abs/2511.15622
• Depth Anything 3 for monocular depth estimation: https://depth-anything-3.github.io/assets/da3_tech_report_2025.pdf

Installation

Conda (recommended)

conda create -n dap-3_py3-11 "python>=3.7,<3.11"
conda activate dap-3_py3-11
pip install torch>=2 torchvision --index-url https://download.pytorch.org/whl/cu126
pip install xformers
git clone https://github.com/obrookes/Unmarked-Anything
cd Unmarked-Anything
pip install -e .
pip install gsplat
pip install ultralytics

For AArch/ARM HPC environments where pycolmap is unavailable, keep pycolmap disabled in pyproject.toml (currently commented out). For the same environment, pycolmap import is also commented in src/depth_anything_3/utils/export/colmap.py, so COLMAP export paths that depend on pycolmap are disabled.

Singularity Container

Container assets:

• Definition: envs/containers/depth-anything-3.def
• Local image: envs/containers/dap3-ultralytics.sif (git-ignored)

Build image (if needed):

singularity build envs/containers/dap3-ultralytics.sif envs/containers/depth-anything-3.def

Run CLI in container:

singularity exec --nv envs/containers/dap3-ultralytics.sif \
  python apps/camera_trap/cli/dap3_cli.py \
  --input-video-dir assets/videos \
  --output-dir outputs/demo \
  --sam3-model-path weights/sam3/safari_checkpoint_hf.pt \
  --sam3-text-prompts animal

Use --nv when you need GPU support in the container.

Model Files

Store large local model files here (all git-ignored):

• weights/sam3/ for SAM3 checkpoints
• weights/dap3/ for DA3 checkpoints/artifacts

Current SAM3 checkpoint example:

• weights/sam3/safari_checkpoint_hf.pt

Quick Start

Canonical CLI:

python apps/camera_trap/cli/dap3_cli.py \
  --input-video-dir assets/videos \
  --output-dir outputs/demo \
  --sam3-model-path weights/sam3/safari_checkpoint_hf.pt \
  --sam3-text-prompts animal

Legacy wrapper also works: python dap3_cli.py ...

Sample Data

• Input video: assets/videos/03290392_0_10.MP4
• Suggested output directory: outputs/demo/
• Visualization notebook: notebooks/camera_trap/visualize_test_output.ipynb (configured for assets/videos and outputs/demo; includes a DAP3-style selected-frame analysis cell with manual/auto frame selection, depth-map mask contour overlay, bbox/center annotations, and mask-vs-bbox depth statistics/distributions)

HPC Execution

• For SLURM/HPC job structure, batch scripts, manifests, and helper utilities, see: hpc/README.md

Full Command Template

python apps/camera_trap/cli/dap3_cli.py \
  --input-video-dir /data/camera_trap/videos \
  --output-dir /data/camera_trap/results \
  --sam3-model-path weights/sam3/safari_checkpoint_hf.pt \
  --sam3-text-prompts animal deer boar \
  --da3-model-id depth-anything/DA3NESTED-GIANT-LARGE \
  --target-fps 1.0 \
  --sam3-mode track \
  --conf 0.25 \
  --device auto

CLI Options (All)

Option	Type	Required	Default	Description
--input-video-dir	path	yes	-	Directory containing videos to process.
--output-dir	path	yes	-	Output root for run manifest and per-video results.
--video-exts	string	no	.mp4,.mov,.avi,.mkv	Comma-separated extensions (with or without leading .).
--sam3-model-path	path	yes	-	Path to SAM3 checkpoint (.pt).
--sam3-text-prompts	list of strings	yes	-	One or more global SAM3 text prompts.
--da3-model-id	string	no	depth-anything/DA3NESTED-GIANT-LARGE	DA3 pretrained model ID.
--target-fps	float	no	1.0	Sampling rate for processing. Must be > 0.
--sam3-mode	enum	no	track	track (video tracking) or frame (per-frame segmentation).
--conf	float	no	0.25	SAM3 confidence threshold.
--device	enum	no	auto	auto, cuda, or cpu.
--half	flag	no	false	Enable FP16 for SAM3 (CUDA only).
--overwrite	flag	no	false	Reprocess even if output JSON + NPZ already exist.
--max-videos	int	no	None	Cap number of videos after sorting. Must be > 0 if provided.

Output Structure

Given --output-dir ./out, outputs are:

• ./out/run_manifest.json
• ./out/<video_stem>/<video_stem>.json
• ./out/<video_stem>/<video_stem>_arrays.npz

Per-frame status values include: processed, empty_mask, sam_error, da3_error, frame_decode_error.

Notes

• target_fps is implemented as frame stride (round(video_fps / target_fps), minimum 1).
• Non-overwrite mode skips videos that already have both expected output files.
• If --sam3-mode track is unavailable in your ultralytics build, use --sam3-mode frame.

Additional References

• Camera trap CLI source: apps/camera_trap/cli/dap3_cli.py
• Archive index: archive/INDEX.md
• Archived upstream README: archive/2026-03/readmes/README.upstream.md