workers/tilt — Single-Image Camera Calibration (roll / pitch / FOV)
Predicts camera roll, pitch, and vertical field-of-view from one image, based on AlanSavio25/DeepSingleImageCalibration.
- Dockerfile:
/workers/tilt/Dockerfile(build context = repo root) - GPU: light — NVIDIA T4 (16 GB) is plenty; CPU also works, slower.
- App env var to set on the endpoint:
RUNPOD_TILT_URL— the deployed RunPod endpoint URL the app calls.
Contract
Input:
{ "input": { "image": "<base64>" } }
image may include a data:image/...;base64, prefix (stripped automatically).
Output:
{ "roll_degrees": 0.0, "pitch_degrees": 0.0, "fov_degrees": 0.0 }
On error: { "error": "..." }.
Weights (required — research repo, not on pip)
The pretrained checkpoint is not bundled. Download the repo's released weights and place them on the network volume, then point the worker at them:
- Default path:
/runpod-volume/tilt/weights.tar - Override with env var
WEIGHTS_PATH.
The backbone loads into the volume-cached TORCH_HOME/HF_HOME on first request.
Tunable env vars (no rebuild needed)
WEIGHTS_PATH, CALIB_BACKBONE (densenet161|densenet121), CALIB_NUM_BINS,
CALIB_INPUT_SIZE, CALIB_ROLL_MIN/MAX, CALIB_VFOV_MIN/MAX, CALIB_RHO_MIN/MAX,
CALIB_DIRECT_PITCH (=1 if the net outputs pitch directly), CALIB_PITCH_MIN/MAX.