Doni/safesight
6
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

o4.6做了修改

Browse Source
This commit is contained in:
sladro 2026-02-27 22:03:28 +08:00
parent aaa794fe10
commit 631f2a1f4a
3 changed files with 39 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
configs/test_cam1_ppe11_yolo_debug.json
View File

@ -35,7 +35,7 @@
"dst_h": 768,
"dst_format": "rgb",
"dst_packed": true,
"resize_mode": "letterbox",
"resize_mode": "stretch",
"keep_ratio": false,
"rga_gate": "cam_ppe11_yolo_debug",
"use_rga": true

24
plugins/ai_yolo/ai_yolo_node.cpp
View File

@ -951,11 +951,11 @@ private:
NMS(valid_count, boxes, class_ids, indices, c, nms_thresh_);
}
const DetCoordContext coord_ctx = BuildDetCoordContext(*frame, model_input_w_, model_input_h_);
auto det_result = std::make_shared<DetectionResult>();
det_result->img_w = coord_ctx.out_w;
det_result->img_h = coord_ctx.out_h;
// README 约定: img_w/img_h = 推理输入宽/高（当前帧空间）
// 坐标缩放由下游 preprocess 的 ScaleMeta 负责
det_result->img_w = frame->width;
det_result->img_h = frame->height;
det_result->model_name = (yolo_version_ == YoloVersion::V5) ? "yolov5" : "yolov8";
for (int i = 0; i < valid_count && det_result->items.size() < kMaxDetections; ++i) {
@ -975,7 +975,10 @@ private:
Detection det;
det.cls_id = cls_id;
det.score = obj_probs[i];
det.bbox = DecodeToOutputRect(x1, y1, w, h, coord_ctx);
det.bbox.x = std::max(0.0f, std::min(x1, static_cast<float>(frame->width)));
det.bbox.y = std::max(0.0f, std::min(y1, static_cast<float>(frame->height)));
det.bbox.w = std::max(0.0f, std::min(w, static_cast<float>(frame->width) - det.bbox.x));
det.bbox.h = std::max(0.0f, std::min(h, static_cast<float>(frame->height) - det.bbox.y));
det.track_id = -1;
if (debug_det_ && det_result->items.size() < 5 && processed_ < 20) {
@ -1101,11 +1104,9 @@ private:
NMS(valid_count, boxes, class_ids, indices, c, nms_thresh_);
}
const DetCoordContext coord_ctx = BuildDetCoordContext(*frame, model_input_w_, model_input_h_);
auto det_result = std::make_shared<DetectionResult>();
det_result->img_w = coord_ctx.out_w;
det_result->img_h = coord_ctx.out_h;
det_result->img_w = frame->width;
det_result->img_h = frame->height;
det_result->model_name = (yolo_version_ == YoloVersion::V5) ? "yolov5" : "yolov8";
for (int i = 0; i < valid_count && det_result->items.size() < kMaxDetections; ++i) {
@ -1125,7 +1126,10 @@ private:
Detection det;
det.cls_id = cls_id;
det.score = obj_probs[i];
det.bbox = DecodeToOutputRect(x1, y1, w, h, coord_ctx);
det.bbox.x = std::max(0.0f, std::min(x1, static_cast<float>(frame->width)));
det.bbox.y = std::max(0.0f, std::min(y1, static_cast<float>(frame->height)));
det.bbox.w = std::max(0.0f, std::min(w, static_cast<float>(frame->width) - det.bbox.x));
det.bbox.h = std::max(0.0f, std::min(h, static_cast<float>(frame->height) - det.bbox.y));
det.track_id = -1;
if (debug_det_ && det_result->items.size() < 5 && processed_ < 20) {

26
plugins/osd/osd_node.cpp
View File

@ -77,6 +77,21 @@ Rect MapRectToFrame(const Rect& in, int src_w, int src_h, int dst_w, int dst_h)
return out;
}
// 利用 transform_meta 将原图空间坐标正确映射回当前帧空间
// 正向变换: frame_x = orig_x * scale_x + pad_x
Rect MapRectWithTransform(const Rect& in, const FrameTransformMeta& tx, int frame_w, int frame_h) {
const float fw = static_cast<float>(frame_w);
const float fh = static_cast<float>(frame_h);
Rect out{};
out.x = std::max(0.0f, in.x * tx.scale_x + tx.pad_x);
out.y = std::max(0.0f, in.y * tx.scale_y + tx.pad_y);
out.w = std::max(0.0f, in.w * tx.scale_x);
out.h = std::max(0.0f, in.h * tx.scale_y);
if (out.x + out.w > fw) out.w = std::max(0.0f, fw - out.x);
if (out.y + out.h > fh) out.h = std::max(0.0f, fh - out.y);
return out;
}
#if defined(RK3588_ENABLE_RGA)
inline uint32_t PackColorArgb(const Color& c) {
return (0xFFu << 24) | (static_cast<uint32_t>(c.r) << 16) |
@ -578,6 +593,11 @@ private:
const int det_w = frame->det->img_w > 0 ? frame->det->img_w : w;
const int det_h = frame->det->img_h > 0 ? frame->det->img_h : h;
const bool map_det_to_frame = (det_w != w || det_h != h);
// 优先使用 transform_meta 做 letterbox 感知的坐标映射
const bool use_transform = map_det_to_frame && frame->transform_meta &&
frame->transform_meta->valid &&
frame->transform_meta->scale_x > 1e-6f &&
frame->transform_meta->scale_y > 1e-6f;
uint8_t* data = frame->planes[0].data ? frame->planes[0].data : frame->data;
PixelFormat fmt = frame->format;
@ -604,7 +624,8 @@ private:
} else {
bool ok = true;
for (const auto& det : frame->det->items) {
const Rect draw = map_det_to_frame ? MapRectToFrame(det.bbox, det_w, det_h, w, h) : det.bbox;
const Rect draw = use_transform ? MapRectWithTransform(det.bbox, *frame->transform_meta, w, h)
: (map_det_to_frame ? MapRectToFrame(det.bbox, det_w, det_h, w, h) : det.bbox);
int x = Clamp(static_cast<int>(draw.x), 0, w - 1);
int y = Clamp(static_cast<int>(draw.y), 0, h - 1);
int rw = static_cast<int>(draw.w);
@ -654,7 +675,8 @@ private:
}
for (const auto& det : frame->det->items) {
const Rect draw = map_det_to_frame ? MapRectToFrame(det.bbox, det_w, det_h, w, h) : det.bbox;
const Rect draw = use_transform ? MapRectWithTransform(det.bbox, *frame->transform_meta, w, h)
: (map_det_to_frame ? MapRectToFrame(det.bbox, det_w, det_h, w, h) : det.bbox);
int x1 = static_cast<int>(draw.x);
int y1 = static_cast<int>(draw.y);
int x2 = static_cast<int>(draw.x + draw.w);