36 lines
1.4 KiB
Markdown
36 lines
1.4 KiB
Markdown
# 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](https://github.com/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:
|
|
```json
|
|
{ "input": { "image": "<base64>" } }
|
|
```
|
|
`image` may include a `data:image/...;base64,` prefix (stripped automatically).
|
|
|
|
Output:
|
|
```json
|
|
{ "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`.
|