luboslenco
56 lines
1.6 KiB
C
56 lines
1.6 KiB
C
/*
|
|
* iris_depth.h - Depth Anything 3 (DA3MONO-LARGE) monocular depth estimation
|
|
*
|
|
* Self-contained inference for the Depth Anything 3 "mono" model
|
|
* (DA3MONO-LARGE): a DINOv2 ViT-L/14 backbone feeding a DPT-style head.
|
|
* Loads weights from a single safetensors file (all tensors F32) and
|
|
* produces a single-channel relative-depth map for an RGB image.
|
|
*
|
|
* The heavy transformer matrix multiplies are offloaded to the Vulkan
|
|
* GEMM backend when it is available; otherwise everything runs on the CPU.
|
|
*/
|
|
|
|
#ifndef IRIS_DEPTH_H
|
|
#define IRIS_DEPTH_H
|
|
|
|
#include "iris.h"
|
|
|
|
#ifdef __cplusplus
|
|
extern "C" {
|
|
#endif
|
|
|
|
typedef struct iris_depth iris_depth_t;
|
|
|
|
/*
|
|
* Load the Depth Anything V2 model from a safetensors file.
|
|
* Returns NULL on error (use iris_get_error() for details).
|
|
*/
|
|
iris_depth_t *iris_depth_load(const char *path);
|
|
|
|
/*
|
|
* Enable "tileable" mode: remove the perspective front-to-back brightness
|
|
* ramp the model bakes into the depth, then feather opposite edges together,
|
|
* so that depth estimated from a seamless top-down texture also tiles
|
|
* seamlessly. Off by default.
|
|
*/
|
|
void iris_depth_set_tileable(iris_depth_t *model, int on);
|
|
|
|
/*
|
|
* Free the model and its resources.
|
|
*/
|
|
void iris_depth_free(iris_depth_t *model);
|
|
|
|
/*
|
|
* Estimate depth for an RGB(A) image. Alpha is ignored. The result is a
|
|
* single-channel (grayscale) image with the same dimensions as the input,
|
|
* where brighter pixels are nearer to the camera (normalized to 0..255).
|
|
* Returns a newly allocated image (free with iris_image_free), NULL on error.
|
|
*/
|
|
iris_image *iris_depth_estimate(iris_depth_t *model, const iris_image *input);
|
|
|
|
#ifdef __cplusplus
|
|
}
|
|
#endif
|
|
|
|
#endif /* IRIS_DEPTH_H */
|