diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-10-21 17:47:57 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-10-21 17:47:57 +0200 |
| commit | 8b8acffcf5f1e9e9ddb552b8ac125e50f68987ac (patch) | |
| tree | 9b618133a468585087ff4b014b22029d40b24bc5 | |
| parent | 2ce585cc96a3fffba65f1bafbfbc796c6370308f (diff) | |
| download | facedet-0.1.0.tar.gz facedet-0.1.0.zip | |
| -rw-r--r-- | Makefile | 18 | ||||
| -rw-r--r-- | README.md | 30 | ||||
| -rw-r--r-- | convert.py | 6 | ||||
| -rw-r--r-- | facedet.c (renamed from yunet.c) | 108 | ||||
| -rw-r--r-- | facedet.h (renamed from yunet.h) | 34 | ||||
| -rw-r--r-- | gemm.c | 10 | ||||
| -rw-r--r-- | gemm.h | 10 | ||||
| -rw-r--r-- | image.h | 4 | ||||
| -rw-r--r-- | main.c | 22 | ||||
| -rw-r--r-- | nn.c | 20 | ||||
| -rw-r--r-- | nn.h | 18 | ||||
| -rw-r--r-- | overlay.c | 4 | ||||
| -rw-r--r-- | overlay.h | 8 |
13 files changed, 146 insertions, 146 deletions
@@ -4,20 +4,20 @@ CFLAGS = -O3 -std=c11 -Wall -Wextra -Wpedantic LDLIBS = -lm BUILD = build -CORE = yunet.c nn.c gemm.c +CORE = facedet.c nn.c gemm.c MODEL_HEADER = model_data.h -MODEL_CPPFLAGS = -DYUNET_MODEL_HEADER='"$(abspath $(MODEL_HEADER))"' +MODEL_CPPFLAGS = -DFACEDET_MODEL_HEADER='"$(abspath $(MODEL_HEADER))"' REFINER_HEADER = ifneq ($(strip $(REFINER_HEADER)),) -MODEL_CPPFLAGS += -DYUNET_REFINER_HEADER='"$(abspath $(REFINER_HEADER))"' +MODEL_CPPFLAGS += -DFACEDET_REFINER_HEADER='"$(abspath $(REFINER_HEADER))"' endif -HEADERS = yunet.h nn.h gemm.h $(MODEL_HEADER) $(REFINER_HEADER) -OBJECTS = $(BUILD)/yunet.o $(BUILD)/nn.o $(BUILD)/gemm.o +HEADERS = facedet.h nn.h gemm.h $(MODEL_HEADER) $(REFINER_HEADER) +OBJECTS = $(BUILD)/facedet.o $(BUILD)/nn.o $(BUILD)/gemm.o .PHONY: all model clean -all: $(BUILD)/yunet $(BUILD)/libyunet.a $(BUILD)/libyunet.so $(BUILD)/libimage.so +all: $(BUILD)/facedet $(BUILD)/libfacedet.a $(BUILD)/libfacedet.so $(BUILD)/libimage.so $(BUILD): mkdir -p $(BUILD) @@ -25,7 +25,7 @@ $(BUILD): $(BUILD)/%.o: %.c $(HEADERS) | $(BUILD) $(CC) $(CPPFLAGS) $(MODEL_CPPFLAGS) $(CFLAGS) -c $< -o $@ -$(BUILD)/libyunet.a: $(OBJECTS) +$(BUILD)/libfacedet.a: $(OBJECTS) $(AR) rcs $@ $^ $(BUILD)/main.o $(BUILD)/image.o: image.h @@ -33,10 +33,10 @@ $(BUILD)/image.o: third_party/stb/stb_image.h $(BUILD)/main.o $(BUILD)/overlay.o: overlay.h $(BUILD)/overlay.o $(BUILD)/draw.o: draw.h -$(BUILD)/yunet: $(BUILD)/main.o $(BUILD)/image.o $(BUILD)/overlay.o $(BUILD)/draw.o $(BUILD)/libyunet.a +$(BUILD)/facedet: $(BUILD)/main.o $(BUILD)/image.o $(BUILD)/overlay.o $(BUILD)/draw.o $(BUILD)/libfacedet.a $(CC) $(CFLAGS) $(LDFLAGS) -o $@ $^ $(LDLIBS) -$(BUILD)/libyunet.so: $(CORE) $(HEADERS) | $(BUILD) +$(BUILD)/libfacedet.so: $(CORE) $(HEADERS) | $(BUILD) $(CC) $(CPPFLAGS) $(MODEL_CPPFLAGS) $(CFLAGS) $(LDFLAGS) -shared -fPIC -o $@ $(CORE) $(LDLIBS) $(BUILD)/libimage.so: image.c image.h third_party/stb/stb_image.h | $(BUILD) @@ -1,6 +1,6 @@ -# YuNet 2026may +# facedet -A CPU-only face detector for `face_detection_yunet_2026may`. +A CPU-only face detector using the YuNet `face_detection_yunet_2026may` model. ## Build and run @@ -11,11 +11,11 @@ embedded in the binary. ```sh make -./build/yunet photo.jpg -./build/yunet photo.png 0.6 0.3 -./build/yunet image.ppm -./build/yunet photo.jpg -o overlay.ppm > faces.json -./build/yunet --out overlay.ppm photo.png 0.6 0.3 +./build/facedet photo.jpg +./build/facedet photo.png 0.6 0.3 +./build/facedet image.ppm +./build/facedet photo.jpg -o overlay.ppm > faces.json +./build/facedet --out overlay.ppm photo.png 0.6 0.3 ``` @@ -26,7 +26,7 @@ Input is JPEG, PNG, or binary PPM (`P6`, RGB, maxval 255), with dimensions from 1 to 4096. ```sh -./build/yunet photo.jpg > faces.json +./build/facedet photo.jpg > faces.json ``` Standard output contains one JSON object: @@ -51,26 +51,26 @@ decoded image with green face boxes, confidence labels, and five landmark dots. ## C API -Include `yunet.h` and link `build/libyunet.a` with `-lm`. +Include `facedet.h` and link `build/libfacedet.a` with `-lm`. -1. Create a context with `yunet_create(padded_width, padded_height)`. Both dimensions +1. Create a context with `facedet_create(padded_width, padded_height)`. Both dimensions must be multiples of 32 in the range 32 through 4096. Check for `NULL`. -2. Fill it with `yunet_set_rgb`, or fill the buffer returned by `yunet_input` with +2. Fill it with `facedet_set_rgb`, or fill the buffer returned by `facedet_input` with float32 BGR CHW values in `[0, 255]`. -3. Call `yunet_run`, then `yunet_decode` with an output array and its capacity. +3. Call `facedet_run`, then `facedet_decode` with an output array and its capacity. Negative return values indicate errors. The output capacity limits returned detections, not the number of candidates considered. -4. Reuse the context for images with the same padded dimensions. Call `yunet_free` +4. Reuse the context for images with the same padded dimensions. Call `facedet_free` when finished. Use separate contexts for concurrent inference. The command-line implementation in `main.c` demonstrates the complete flow. -`yunet_map` exposes raw CHW head tensors for numerical checks; classification and +`facedet_map` exposes raw CHW head tensors for numerical checks; classification and objectness maps contain logits, so the decoder applies sigmoid before computing `sqrt(classification * objectness)`. `nn.c` contains the required convolution, pooling, addition, and resizing operations. `gemm.c` uses AArch64 NEON or runtime-selected x86-64 AVX2/FMA, -with a portable C fallback. Define `YUNET_SCALAR` to disable explicit SIMD kernels. +with a portable C fallback. Define `FACEDET_SCALAR` to disable explicit SIMD kernels. ## Model and regeneration @@ -153,7 +153,7 @@ def main(): onnx.checker.check_model(model) heads = head_map(nodes) plan, weights, shapes, offsets, arena = compile_plan(nodes, init, heads) - lines = ["#ifndef YUNET_MODEL_DATA_H", "#define YUNET_MODEL_DATA_H", "", + lines = ["#ifndef FACEDET_MODEL_DATA_H", "#define FACEDET_MODEL_DATA_H", "", "static const float model_weight_data[] = {"] weight_offsets = [] pos = 0 @@ -166,14 +166,14 @@ def main(): pos += len(flat) lines += ["};", "", "static const float *const model_weights[] = {"] lines += [f" model_weight_data + {off}," for off in weight_offsets] - lines += ["};", "", "static const yunet_op model_ops[] = {"] + lines += ["};", "", "static const facedet_op model_ops[] = {"] for p in plan: fields = [p["op"], p.get("group", 0), p.get("k", 0), p.get("s", 0), p.get("pad", 0), p.get("relu", 0), p.get("widx", 0), *p["in_shape"], *p["out_shape"], p["in_off"][0], p["in_off"][1] if len(p["in_off"]) > 1 else 0, p["out_off"]] lines.append(" {" + ", ".join(map(str, fields)) + "},") - lines += ["};", "", "static const yunet_map_info model_maps[] = {"] + lines += ["};", "", "static const facedet_map_info model_maps[] = {"] for name, (tensor, sigmoid) in heads.items(): c, h, w = shapes[tensor] lines.append(f' {{"{name}", {offsets[tensor]}, {c}, {h}, {w}, {int(sigmoid)}}},') @@ -1,4 +1,4 @@ -#include "yunet.h" +#include "facedet.h" #include "gemm.h" #include "nn.h" @@ -14,45 +14,45 @@ typedef struct { int op, group, k, s, pad, relu, widx; int cin, ih, iw, cout, oh, ow; size_t in0_off, in1_off, out_off; -} yunet_op; +} facedet_op; typedef struct { const char *name; size_t off; int c, h, w, sigmoid; -} yunet_map_info; +} facedet_map_info; -#ifndef YUNET_MODEL_HEADER -#define YUNET_MODEL_HEADER "model_data.h" +#ifndef FACEDET_MODEL_HEADER +#define FACEDET_MODEL_HEADER "model_data.h" #endif -#include YUNET_MODEL_HEADER -#ifdef YUNET_REFINER_HEADER -#include YUNET_REFINER_HEADER +#include FACEDET_MODEL_HEADER +#ifdef FACEDET_REFINER_HEADER +#include FACEDET_REFINER_HEADER _Static_assert(REFINER_OP_COUNT == MODEL_OP_COUNT, "Refiner operation count mismatch"); _Static_assert(REFINER_MAP_COUNT == MODEL_MAP_COUNT, "Refiner map count mismatch"); _Static_assert(REFINER_ARENA_BYTES == MODEL_ARENA_BYTES, "Refiner arena mismatch"); #endif -struct yunet { +struct facedet { int width, height, has_result; unsigned char *arena; - yunet_op ops[MODEL_OP_COUNT]; - yunet_map_info maps[MODEL_MAP_COUNT]; + facedet_op ops[MODEL_OP_COUNT]; + facedet_map_info maps[MODEL_MAP_COUNT]; float *packs[MODEL_OP_COUNT]; float *im2col, *bscratch; const float *const *weights; int is_refiner; }; -static yunet *create_model(int width, int height, const yunet_op *ops, - const yunet_map_info *maps, const float *const *weights) { +static facedet *create_model(int width, int height, const facedet_op *ops, + const facedet_map_info *maps, const float *const *weights) { if (width < 32 || height < 32 || width > 4096 || height > 4096 || width % 32 || height % 32) return NULL; size_t area = (size_t)(width / 32) * (height / 32); if (area > SIZE_MAX / MODEL_ARENA_BYTES) return NULL; - yunet *m = calloc(1, sizeof *m); + facedet *m = calloc(1, sizeof *m); if (!m) return NULL; m->width = width; @@ -64,7 +64,7 @@ static yunet *create_model(int width, int height, const yunet_op *ops, memset(m->arena, 0, (size_t)width * height * 3 * sizeof(float)); size_t max_col = 0, max_k = 0; for (int i = 0; i < MODEL_OP_COUNT; i++) { - yunet_op op = ops[i]; + facedet_op op = ops[i]; /* All spatial dimensions scale from the 32x32 plan. Scaling every * arena slot by the same area preserves its capacity and lifetime. */ op.ih *= height / 32; @@ -77,10 +77,10 @@ static yunet *create_model(int width, int height, const yunet_op *ops, m->ops[i] = op; if (op.op == OP_CONV && op.group == 1) { int k = op.cin * op.k * op.k; - m->packs[i] = malloc(yunet_gemm_pack_a_bytes(op.cout, k)); + m->packs[i] = malloc(facedet_gemm_pack_a_bytes(op.cout, k)); if (!m->packs[i]) goto fail; - yunet_gemm_pack_a(weights[op.widx], op.cout, k, m->packs[i]); + facedet_gemm_pack_a(weights[op.widx], op.cout, k, m->packs[i]); if ((size_t)k > max_k) max_k = (size_t)k; size_t col = op.k == 3 ? (size_t)k * op.oh * op.ow : 0; @@ -102,17 +102,17 @@ static yunet *create_model(int width, int height, const yunet_op *ops, } return m; fail: - yunet_free(m); + facedet_free(m); return NULL; } -yunet *yunet_create(int width, int height) { +facedet *facedet_create(int width, int height) { return create_model(width, height, model_ops, model_maps, model_weights); } -yunet *yunet_create_refiner(void) { -#ifdef YUNET_REFINER_HEADER - yunet *m = create_model(160, 160, refiner_ops, refiner_maps, refiner_weights); +facedet *facedet_create_refiner(void) { +#ifdef FACEDET_REFINER_HEADER + facedet *m = create_model(160, 160, refiner_ops, refiner_maps, refiner_weights); if (m) m->is_refiner = 1; return m; @@ -121,7 +121,7 @@ yunet *yunet_create_refiner(void) { #endif } -void yunet_free(yunet *m) { +void facedet_free(facedet *m) { if (!m) return; for (int i = 0; i < MODEL_OP_COUNT; i++) @@ -132,19 +132,19 @@ void yunet_free(yunet *m) { free(m); } -float *yunet_input(yunet *m) { +float *facedet_input(facedet *m) { if (!m) return NULL; m->has_result = 0; return (float *)m->arena; } -int yunet_set_rgb(yunet *m, const unsigned char *rgb, int width, int height, size_t row_bytes) { +int facedet_set_rgb(facedet *m, const unsigned char *rgb, int width, int height, size_t row_bytes) { if (!m || !rgb || width < 1 || height < 1 || width > m->width || height > m->height || row_bytes < (size_t)width * 3 || (height > 1 && row_bytes > (SIZE_MAX - (size_t)width * 3) / (size_t)(height - 1))) return -1; - float *input = yunet_input(m); + float *input = facedet_input(m); size_t plane = (size_t)m->width * m->height; memset(input, 0, plane * 3 * sizeof(float)); for (int y = 0; y < height; y++) { @@ -156,7 +156,7 @@ int yunet_set_rgb(yunet *m, const unsigned char *rgb, int width, int height, siz return 0; } -int yunet_run(yunet *m) { +int facedet_run(facedet *m) { if (!m) return -1; m->has_result = 0; @@ -166,7 +166,7 @@ int yunet_run(yunet *m) { return -1; unsigned char *a = m->arena; for (int i = 0; i < MODEL_OP_COUNT; i++) { - yunet_op op = m->ops[i]; + facedet_op op = m->ops[i]; const float *in0 = (const float *)(a + op.in0_off); float *out = (float *)(a + op.out_off); switch (op.op) { @@ -174,29 +174,29 @@ int yunet_run(yunet *m) { size_t n = (size_t)op.oh * op.ow; const float *bias = m->weights[op.widx + 1]; if (op.group == 1 && op.k == 1 && op.s == 1) { - yunet_gemm(m->packs[i], op.cout, op.cin, in0, (size_t)op.ih * op.iw, (int)n, + facedet_gemm(m->packs[i], op.cout, op.cin, in0, (size_t)op.ih * op.iw, (int)n, bias, op.relu, out, n, m->bscratch); } else if (op.group == 1 && op.k == 3) { - yunet_nn_im2col3(in0, op.cin, op.ih, op.iw, op.s, m->im2col, op.oh, op.ow); - yunet_gemm(m->packs[i], op.cout, op.cin * 9, m->im2col, n, (int)n, + facedet_nn_im2col3(in0, op.cin, op.ih, op.iw, op.s, m->im2col, op.oh, op.ow); + facedet_gemm(m->packs[i], op.cout, op.cin * 9, m->im2col, n, (int)n, bias, op.relu, out, n, m->bscratch); } else if (op.group == op.cin && op.group == op.cout && op.k == 3 && op.s == 1) { - yunet_nn_dwconv3(in0, op.cin, op.ih, op.iw, m->weights[op.widx], bias, + facedet_nn_dwconv3(in0, op.cin, op.ih, op.iw, m->weights[op.widx], bias, op.relu, out); } else { - yunet_nn_conv2d(in0, op.cin, op.ih, op.iw, m->weights[op.widx], bias, + facedet_nn_conv2d(in0, op.cin, op.ih, op.iw, m->weights[op.widx], bias, op.cout, op.k, op.s, op.pad, op.group, op.relu, out, op.oh, op.ow); } break; } case OP_MAXPOOL: - yunet_nn_maxpool2(in0, op.cin, op.ih, op.iw, out); + facedet_nn_maxpool2(in0, op.cin, op.ih, op.iw, out); break; case OP_RESIZE2X: - yunet_nn_resize2x(in0, op.cin, op.ih, op.iw, out); + facedet_nn_resize2x(in0, op.cin, op.ih, op.iw, out); break; case OP_ADD: - yunet_nn_add(in0, (const float *)(a + op.in1_off), out, (size_t)op.cin * op.ih * op.iw); + facedet_nn_add(in0, (const float *)(a + op.in1_off), out, (size_t)op.cin * op.ih * op.iw); break; } } @@ -204,11 +204,11 @@ int yunet_run(yunet *m) { return 0; } -const float *yunet_map(const yunet *m, const char *name, int *channels, int *height, int *width) { +const float *facedet_map(const facedet *m, const char *name, int *channels, int *height, int *width) { if (!m || !name || !m->has_result) return NULL; for (int i = 0; i < MODEL_MAP_COUNT; i++) { - const yunet_map_info *mm = &m->maps[i]; + const facedet_map_info *mm = &m->maps[i]; if (strcmp(mm->name, name) == 0) { if (channels) *channels = mm->c; if (height) *height = mm->h; @@ -239,8 +239,8 @@ static float sample_rgb(const unsigned char *rgb, int width, int height, size_t return (float)value; } -int yunet_refine(yunet *m, const unsigned char *rgb, int width, int height, - size_t row_bytes, yunet_detection *faces, int count) { +int facedet_refine(facedet *m, const unsigned char *rgb, int width, int height, + size_t row_bytes, facedet_detection *faces, int count) { if (!m || !m->is_refiner || !rgb || width < 1 || height < 1 || width > 4096 || height > 4096 || count < 0 || (!faces && count) || row_bytes < (size_t)width * 3 || @@ -254,14 +254,14 @@ int yunet_refine(yunet *m, const unsigned char *rgb, int width, int height, if (!points) return -1; for (int i = 0; i < count; i++) { - const yunet_detection *face = &faces[i]; + const facedet_detection *face = &faces[i]; if (!isfinite(face->x) || !isfinite(face->y) || !isfinite(face->w) || !isfinite(face->h) || face->w <= 0 || face->h <= 0) goto fail; double extent = fmax(face->w, face->h) * 1.3; double ox = face->x + (double)face->w / 2 - extent / 2; double oy = face->y + (double)face->h / 2 - extent / 2; - float *input = yunet_input(m); + float *input = facedet_input(m); for (int y = 0; y < 160; y++) for (int x = 0; x < 160; x++) for (int c = 0; c < 3; c++) @@ -269,11 +269,11 @@ int yunet_refine(yunet *m, const unsigned char *rgb, int width, int height, sample_rgb(rgb, width, height, row_bytes, ox + (x + .5) * extent / 160 - .5, oy + (y + .5) * extent / 160 - .5, 2 - c); - if (yunet_run(m)) + if (facedet_run(m)) goto fail; int channels, h, w; /* The landmark readout is the fixed center cell, independent of objectness. */ - const float *map = yunet_map(m, "kps_8", &channels, &h, &w); + const float *map = facedet_map(m, "kps_8", &channels, &h, &w); if (!map || channels != 10 || h != 20 || w != 20) goto fail; for (int k = 0; k < 5; k++) { @@ -294,7 +294,7 @@ fail: return -1; } -static float iou_xywh(const yunet_detection *a, const yunet_detection *b) { +static float iou_xywh(const facedet_detection *a, const facedet_detection *b) { float x0 = a->x > b->x ? a->x : b->x; float y0 = a->y > b->y ? a->y : b->y; float x1 = (a->x + a->w) < (b->x + b->w) ? a->x + a->w : b->x + b->w; @@ -307,7 +307,7 @@ static float iou_xywh(const yunet_detection *a, const yunet_detection *b) { } static int det_cmp(const void *pa, const void *pb) { - const yunet_detection *a = pa, *b = pb; + const facedet_detection *a = pa, *b = pb; if (a->score != b->score) return a->score > b->score ? -1 : 1; if (a->y != b->y) @@ -315,7 +315,7 @@ static int det_cmp(const void *pa, const void *pb) { return a->x < b->x ? -1 : a->x > b->x; } -int yunet_decode(const yunet *m, float score_thresh, float nms_iou, yunet_detection *out, int cap) { +int facedet_decode(const facedet *m, float score_thresh, float nms_iou, facedet_detection *out, int cap) { if (!m || m->is_refiner || !m->has_result || (!out && cap > 0) || cap < 0 || !isfinite(score_thresh) || score_thresh < 0.0f || score_thresh > 1.0f || !isfinite(nms_iou) || nms_iou < 0.0f || nms_iou > 1.0f) @@ -331,24 +331,24 @@ int yunet_decode(const yunet *m, float score_thresh, float nms_iou, yunet_detect size_t count = 0; for (int lv = 0; lv < 3; lv++) count += (size_t)(m->height / strides[lv]) * (m->width / strides[lv]); - yunet_detection *cand = malloc(sizeof(*cand) * count); + facedet_detection *cand = malloc(sizeof(*cand) * count); if (!cand) return -1; int n = 0; for (int lv = 0; lv < 3; lv++) { int H, W, s = strides[lv]; - const float *cls = yunet_map(m, names[lv][0], NULL, &H, &W); - const float *obj = yunet_map(m, names[lv][1], NULL, NULL, NULL); - const float *bbox = yunet_map(m, names[lv][2], NULL, NULL, NULL); - const float *kps = yunet_map(m, names[lv][3], NULL, NULL, NULL); + const float *cls = facedet_map(m, names[lv][0], NULL, &H, &W); + const float *obj = facedet_map(m, names[lv][1], NULL, NULL, NULL); + const float *bbox = facedet_map(m, names[lv][2], NULL, NULL, NULL); + const float *kps = facedet_map(m, names[lv][3], NULL, NULL, NULL); size_t hw = (size_t)H * W; for (int y = 0; y < H; y++) { for (int x = 0; x < W; x++) { size_t i = (size_t)y * W + x; - float sc = sqrtf(yunet_nn_sigmoid(cls[i]) * yunet_nn_sigmoid(obj[i])); + float sc = sqrtf(facedet_nn_sigmoid(cls[i]) * facedet_nn_sigmoid(obj[i])); if (!isfinite(sc) || sc < score_thresh) continue; - yunet_detection *d = &cand[n]; + facedet_detection *d = &cand[n]; float cx = (x + bbox[i]) * s; float cy = (y + bbox[hw + i]) * s; d->w = expf(bbox[2 * hw + i]) * s; @@ -1,55 +1,55 @@ -#ifndef YUNET_H -#define YUNET_H +#ifndef FACEDET_H +#define FACEDET_H #include <stddef.h> -typedef struct yunet yunet; +typedef struct facedet facedet; typedef struct { float x, y, w, h; float lmk[5][2]; float score; -} yunet_detection; +} facedet_detection; /* Each dimension must be a multiple of 32 between 32 and 4096. * A context owns mutable scratch space and is not safe for concurrent calls. */ -yunet *yunet_create(int width, int height); -void yunet_free(yunet *m); +facedet *facedet_create(int width, int height); +void facedet_free(facedet *m); /* The input is float32 BGR in CHW order, with values in [0, 255]. * The returned buffer belongs to the context. */ -float *yunet_input(yunet *m); +float *facedet_input(facedet *m); /* Interleaved RGB8 is converted to BGR CHW without resizing. The right and * bottom padding are zero. Width and height cannot exceed the context size. * The RGB buffer must contain at least (height-1)*row_bytes + width*3 bytes. */ -int yunet_set_rgb(yunet *m, const unsigned char *rgb, int width, int height, size_t row_bytes); -int yunet_run(yunet *m); +int facedet_set_rgb(facedet *m, const unsigned char *rgb, int width, int height, size_t row_bytes); +int facedet_run(facedet *m); /* Names are cls_s, obj_s, bbox_s, kps_s, with s in {8,16,32}. * Maps use CHW order; cls and obj contain logits, not probabilities. * The buffer belongs to the context and is valid until the next input change * or run. NULL means that no result or no matching map is available. */ -const float *yunet_map(const yunet *m, const char *name, int *channels, int *height, int *width); +const float *facedet_map(const facedet *m, const char *name, int *channels, int *height, int *width); /* Detector contexts only. Returns the number of detections, or -1 for * invalid arguments/no result. Results are sorted by score after greedy NMS, in input pixel coordinates, * without clipping. Landmarks: right eye, left eye, nose, right/left mouth. * cap limits returned detections, not candidates considered before sorting. */ -int yunet_decode(const yunet *m, float score_threshold, float nms_iou, - yunet_detection *out, int cap); +int facedet_decode(const facedet *m, float score_threshold, float nms_iou, + facedet_detection *out, int cap); /* Returns NULL if optional refinement weights are unavailable or allocation * fails. This context has a fixed 160x160 BGR input for landmark refinement. */ -yunet *yunet_create_refiner(void); +facedet *facedet_create_refiner(void); /* Replaces landmarks only; boxes, scores, and ordering remain unchanged. * Each box defines a centered square crop with side 1.3*max(width,height). * Bilinear sampling uses pixel centers and zero outside the RGB image. - * RGB dimensions are 1..4096; buffer/stride requirements match yunet_set_rgb. + * RGB dimensions are 1..4096; buffer/stride requirements match facedet_set_rgb. * Returns 0 on success, -1 on error. All detections remain unchanged on error. - * m must be a context returned by yunet_create_refiner. */ -int yunet_refine(yunet *m, const unsigned char *rgb, int width, int height, - size_t row_bytes, yunet_detection *faces, int count); + * m must be a context returned by facedet_create_refiner. */ +int facedet_refine(facedet *m, const unsigned char *rgb, int width, int height, + size_t row_bytes, facedet_detection *faces, int count); #endif @@ -3,7 +3,7 @@ #include <stdlib.h> #include <string.h> -void yunet_gemm_pack_a(const float *a, int m, int k, float *pack) { +void facedet_gemm_pack_a(const float *a, int m, int k, float *pack) { int mt = (m + 7) / 8; for (int t = 0; t < mt; t++) { float *p = pack + (size_t)t * 8 * k; @@ -15,9 +15,9 @@ void yunet_gemm_pack_a(const float *a, int m, int k, float *pack) { } } -size_t yunet_gemm_pack_a_bytes(int m, int k) { return (size_t)((m + 7) / 8) * 8 * k * sizeof(float); } +size_t facedet_gemm_pack_a_bytes(int m, int k) { return (size_t)((m + 7) / 8) * 8 * k * sizeof(float); } -#if defined(__aarch64__) && !defined(YUNET_SCALAR) +#if defined(__aarch64__) && !defined(FACEDET_SCALAR) # include <arm_neon.h> static void kernel_8x12(const float *ap, @@ -113,7 +113,7 @@ static void kernel_8x12_scalar(const float *ap, } } -# if defined(__x86_64__) && !defined(YUNET_SCALAR) +# if defined(__x86_64__) && !defined(FACEDET_SCALAR) # include <immintrin.h> __attribute__((target("avx2,fma"))) static void kernel_8x12_avx2(const float *ap, @@ -202,7 +202,7 @@ static void kernel_8x12(const float *ap, # endif #endif -void yunet_gemm(const float *a_pack, +void facedet_gemm(const float *a_pack, int m, int k, const float *b, @@ -1,12 +1,12 @@ -#ifndef YUNET_GEMM_H -#define YUNET_GEMM_H +#ifndef FACEDET_GEMM_H +#define FACEDET_GEMM_H #include <stddef.h> -void yunet_gemm_pack_a(const float *a, int m, int k, float *pack); -size_t yunet_gemm_pack_a_bytes(int m, int k); +void facedet_gemm_pack_a(const float *a, int m, int k, float *pack); +size_t facedet_gemm_pack_a_bytes(int m, int k); -void yunet_gemm(const float *a_pack, +void facedet_gemm(const float *a_pack, int m, int k, const float *b, @@ -1,5 +1,5 @@ -#ifndef YUNET_IMAGE_H -#define YUNET_IMAGE_H +#ifndef FACEDET_IMAGE_H +#define FACEDET_IMAGE_H /* Decodes JPEG or PNG to interleaved RGB8, without alpha, orientation, or * color-profile transforms. Each dimension is limited to 4096 pixels. @@ -1,4 +1,4 @@ -#include "yunet.h" +#include "facedet.h" #include "image.h" #include "overlay.h" @@ -168,14 +168,14 @@ int main(int argc, char **argv) { return 1; } int iw = (width + 31) & ~31, ih = (height + 31) & ~31; - yunet *m = yunet_create(iw, ih); - int rc = m ? yunet_set_rgb(m, rgb, width, height, (size_t)width * 3) : -1; + facedet *m = facedet_create(iw, ih); + int rc = m ? facedet_set_rgb(m, rgb, width, height, (size_t)width * 3) : -1; if (rc == 0) - rc = yunet_run(m); + rc = facedet_run(m); int cap = 21 * (iw / 32) * (ih / 32); - yunet_detection *faces = rc == 0 ? malloc((size_t)cap * sizeof(*faces)) : NULL; - int n = faces ? yunet_decode(m, score, nms, faces, cap) : -1; - yunet_free(m); + facedet_detection *faces = rc == 0 ? malloc((size_t)cap * sizeof(*faces)) : NULL; + int n = faces ? facedet_decode(m, score, nms, faces, cap) : -1; + facedet_free(m); if (n < 0) { fprintf(stderr, "Inference failed; memory allocation or input validation failed.\n"); free(faces); @@ -183,9 +183,9 @@ int main(int argc, char **argv) { return 1; } if (refine) { - yunet *refiner = yunet_create_refiner(); - rc = refiner ? yunet_refine(refiner, rgb, width, height, (size_t)width * 3, faces, n) : -1; - yunet_free(refiner); + facedet *refiner = facedet_create_refiner(); + rc = refiner ? facedet_refine(refiner, rgb, width, height, (size_t)width * 3, faces, n) : -1; + facedet_free(refiner); if (rc) { fputs("Landmark refinement failed; check embedded weights and available memory.\n", stderr); free(faces); @@ -203,7 +203,7 @@ int main(int argc, char **argv) { printf("{\"width\":%d,\"height\":%d,\"input_width\":%d,\"input_height\":%d,\"faces\":[", width, height, iw, ih); for (int i = 0; i < n; i++) { - const yunet_detection *d = &faces[i]; + const facedet_detection *d = &faces[i]; printf("%s{\"box\":[%.7g,%.7g,%.7g,%.7g],\"score\":%.7g,\"landmarks\":[", i ? "," : "", d->x, d->y, d->w, d->h, d->score); for (int k = 0; k < 5; k++) @@ -4,7 +4,7 @@ #include <math.h> #include <string.h> -void yunet_nn_conv2d(const float *x, +void facedet_nn_conv2d(const float *x, int cin, int ih, int iw, @@ -53,7 +53,7 @@ void yunet_nn_conv2d(const float *x, } } -void yunet_nn_im2col3(const float *x, int cin, int ih, int iw, int stride, float *col, int oh, int ow) { +void facedet_nn_im2col3(const float *x, int cin, int ih, int iw, int stride, float *col, int oh, int ow) { size_t n = (size_t)oh * ow; for (int ic = 0; ic < cin; ic++) { const float *xc = x + (size_t)ic * ih * iw; @@ -93,9 +93,9 @@ void yunet_nn_im2col3(const float *x, int cin, int ih, int iw, int stride, float } } -#if defined(__aarch64__) && !defined(YUNET_SCALAR) +#if defined(__aarch64__) && !defined(FACEDET_SCALAR) # include <arm_neon.h> -void yunet_nn_dwconv3( +void facedet_nn_dwconv3( const float *x, int c, int ih, int iw, const float *w, const float *bias, int relu, float *y) { for (int ch = 0; ch < c; ch++) { const float *xc = x + (size_t)ch * ih * iw; @@ -163,13 +163,13 @@ void yunet_nn_dwconv3( } } #else -void yunet_nn_dwconv3( +void facedet_nn_dwconv3( const float *x, int c, int ih, int iw, const float *w, const float *bias, int relu, float *y) { - yunet_nn_conv2d(x, c, ih, iw, w, bias, c, 3, 1, 1, c, relu, y, ih, iw); + facedet_nn_conv2d(x, c, ih, iw, w, bias, c, 3, 1, 1, c, relu, y, ih, iw); } #endif -void yunet_nn_maxpool2(const float *x, int c, int ih, int iw, float *y) { +void facedet_nn_maxpool2(const float *x, int c, int ih, int iw, float *y) { int oh = ih / 2, ow = iw / 2; for (int ch = 0; ch < c; ch++) { const float *xc = x + (size_t)ch * ih * iw; @@ -184,7 +184,7 @@ void yunet_nn_maxpool2(const float *x, int c, int ih, int iw, float *y) { } } -void yunet_nn_resize2x(const float *x, int c, int ih, int iw, float *y) { +void facedet_nn_resize2x(const float *x, int c, int ih, int iw, float *y) { int oh = ih * 2, ow = iw * 2; for (int ch = 0; ch < c; ch++) { const float *xc = x + (size_t)ch * ih * iw; @@ -198,9 +198,9 @@ void yunet_nn_resize2x(const float *x, int c, int ih, int iw, float *y) { } } -void yunet_nn_add(const float *a, const float *b, float *y, size_t n) { +void facedet_nn_add(const float *a, const float *b, float *y, size_t n) { for (size_t i = 0; i < n; i++) y[i] = a[i] + b[i]; } -float yunet_nn_sigmoid(float v) { return 1.0f / (1.0f + expf(-v)); } +float facedet_nn_sigmoid(float v) { return 1.0f / (1.0f + expf(-v)); } @@ -1,9 +1,9 @@ -#ifndef YUNET_NN_H -#define YUNET_NN_H +#ifndef FACEDET_NN_H +#define FACEDET_NN_H #include <stddef.h> -void yunet_nn_conv2d(const float *x, +void facedet_nn_conv2d(const float *x, int cin, int ih, int iw, @@ -18,13 +18,13 @@ void yunet_nn_conv2d(const float *x, float *y, int oh, int ow); -void yunet_nn_im2col3(const float *x, int cin, int ih, int iw, int stride, float *col, int oh, int ow); +void facedet_nn_im2col3(const float *x, int cin, int ih, int iw, int stride, float *col, int oh, int ow); -void yunet_nn_dwconv3( +void facedet_nn_dwconv3( const float *x, int c, int ih, int iw, const float *w, const float *bias, int relu, float *y); -void yunet_nn_maxpool2(const float *x, int c, int ih, int iw, float *y); -void yunet_nn_resize2x(const float *x, int c, int ih, int iw, float *y); -void yunet_nn_add(const float *a, const float *b, float *y, size_t n); -float yunet_nn_sigmoid(float v); +void facedet_nn_maxpool2(const float *x, int c, int ih, int iw, float *y); +void facedet_nn_resize2x(const float *x, int c, int ih, int iw, float *y); +void facedet_nn_add(const float *a, const float *b, float *y, size_t n); +float facedet_nn_sigmoid(float v); #endif @@ -6,7 +6,7 @@ static const vix_rgb GREEN = {80, 220, 80}; -static void draw_face(const vix_canvas *canvas, const yunet_detection *face, int index) { +static void draw_face(const vix_canvas *canvas, const facedet_detection *face, int index) { double left = face->x, top = face->y; double right = left + face->w, bottom = top + face->h; if (!isfinite(left) || !isfinite(top) || !isfinite(right) || !isfinite(bottom) || @@ -37,7 +37,7 @@ static void draw_face(const vix_canvas *canvas, const yunet_detection *face, int } int overlay_write_ppm(const char *path, unsigned char *rgb, int width, int height, - const yunet_detection *faces, int count) { + const facedet_detection *faces, int count) { if (!path || !*path || !rgb || width < 1 || height < 1 || width > 4096 || height > 4096 || count < 0 || (count && !faces)) return -1; @@ -1,12 +1,12 @@ -#ifndef YUNET_OVERLAY_H -#define YUNET_OVERLAY_H +#ifndef FACEDET_OVERLAY_H +#define FACEDET_OVERLAY_H -#include "yunet.h" +#include "facedet.h" /* Draws on RGB8 pixels in place and writes a P6 image at the original size. * Detection coordinates are clipped for drawing without modifying the records. * Returns 0 on success or -1 on invalid arguments or a file write failure. */ int overlay_write_ppm(const char *path, unsigned char *rgb, int width, int height, - const yunet_detection *faces, int count); + const facedet_detection *faces, int count); #endif |
