Perception: prepare code for multi-scale camera object detection #8649
Conversation
KaWaiTsoiBaidu
requested review from
weidezhang,
xiaoxq,
techoe,
yuliangguo and
gchen-apollo
techoe
changed the title
|
Please fix lint error |
KaWaiTsoiBaidu
force-pushed
the
multi-scale_camera_obj_detection
branch
from
07b743d
to
c73a98a
Compare
| @@ -126,21 +129,25 @@ __global__ void get_object_kernel(int n, | |||
| float scale = obj_data[loc_index]; | |||
| float cx = (w + sigmoid_gpu(loc_data[offset_loc + 0])) / width; | |||
| float cy = (h + sigmoid_gpu(loc_data[offset_loc + 1])) / height; | |||
| float hw = exp(loc_data[offset_loc + 2]) * anchor_data[2 * c] / width * 0.5; | |||
| float hw = exp(max(-10.0f, min(loc_data[offset_loc + 2], 5.0f))) * | |||
There was a problem hiding this comment.
-10.0f and 5.0f to constexpr values in header file
|
|
||
| int num_anchor_per_scale = num_anchor; | ||
| if (multi_scale){ | ||
| num_anchor_per_scale /= 3; |
| const float *ori_data = ori_data_vec[i]; | ||
| const float *dim_data = dim_data_vec[i]; | ||
| const float *anchor_data = yolo_blobs.anchor_blob->gpu_data() | ||
| + num_anchor_per_scale * 2 * i; |
There was a problem hiding this comment.
static_cast(num_anchor_per_scale * 2 * i)
There was a problem hiding this comment.
This is an index, so integer should be used here.
| std::vector<float> conf_score(cpu_cls_data + k * num_candidates, | ||
| cpu_cls_data + (k + 1) * num_candidates); | ||
| std::vector<float> conf_score(cpu_cls_data + k * all_scales_num_candidates, | ||
| cpu_cls_data + (k + 1) * all_scales_num_candidates); |
There was a problem hiding this comment.
align indentation of all arguments in the next lanes to the first argument.
e.g.
std::vector conf_score(
. cpu_cls_data + k * all_scales_num_candidates,
cpu_cls_data + (k + 1) * all_scales_num_candidates);
| int obj_size = | ||
| output_height * output_width * static_cast<int>(anchors_.size()) / 2; | ||
| output_height_scale1 * output_width_scale1 * | ||
| static_cast<int>(anchors_.size()) / 2; |
There was a problem hiding this comment.
How about
int obj_size = (output_height_scale1 * output_width_scale1 *
static_cast(anchors_.size())) / 2;
to reduce truncation error?
There was a problem hiding this comment.
here output_height_scale1 and output_width_scale1 are also integers, so maybe it does not matter?
| (output_height_scale1 * output_width_scale1 + | ||
| output_height_scale2 * output_width_scale2 + | ||
| output_height_scale3 * output_width_scale3) * | ||
| static_cast<int>(anchors_.size()) / 2 / 3; |
There was a problem hiding this comment.
/2/3 to one constrexpr float
There was a problem hiding this comment.
Here, they are all integers in the expression. So I created 2 constexpr for both 2 and 3 and used them here as you suggested above instead of a constexpr float for 2/3.
Change code to accept multi-scale camera object detection outputs of neural networks. Preparation work for Yolo-v3 object detection network.