FilesLast commitLast update
2 months ago
2 months ago
23 days ago
23 days ago
1 month ago
23 days ago
1 month ago
17 days ago
README

ROIPooling

English | 简体中文

Applicable Products

Product Support
Ascend 950PR&950DT products √
Atlas A3 products √
Atlas A2 products √
Atlas 200I/500 A2 inference products ×
Atlas inference products ×
Atlas training products ×

Function

  • Operator function: Performs max pooling on the input feature map by ROI (Region of Interest) regions and outputs pooling results of a fixed size, used for tasks such as object detection.

  • Formula:

For each ROI n (rois[n] = [batch_idx, x1, y1, x2, y2]), channel c, and pooling position (ph,pw):

1. ROI coordinate mapping (uses separate h/w scaling factors, roundf rounds to int, and the end coordinates have no +1 offset):

roi_startw=round(x1⋅spatial_scale_w),roi_starth=round(y1⋅spatial_scale_h)\text{roi\_start}_w = \text{round}(x1 \cdot \text{spatial\_scale\_w}), \quad \text{roi\_start}_h = \text{round}(y1 \cdot \text{spatial\_scale\_h})

roi_endw=round(x2⋅spatial_scale_w),roi_endh=round(y2⋅spatial_scale_h)\text{roi\_end}_w = \text{round}(x2 \cdot \text{spatial\_scale\_w}), \quad \text{roi\_end}_h = \text{round}(y2 \cdot \text{spatial\_scale\_h})

All the preceding coordinates are integers (int type); the +1 offset is on the ROI size (see below), not on the coordinates.

2. ROI size (int arithmetic, +1 on width/height, malformed ROIs are forced to be non-empty):

roi_w=max⁡(roi_endw−roi_startw+1,1),roi_h=max⁡(roi_endh−roi_starth+1,1)\text{roi\_w} = \max(\text{roi\_end}_w - \text{roi\_start}_w + 1, 1), \quad \text{roi\_h} = \max(\text{roi\_end}_h - \text{roi\_start}_h + 1, 1)

3. Bin splitting and boundary clipping (bin_size is float, based on the int roi_w/roi_h):

bin_sizew=roi_wpooled_w,bin_sizeh=roi_hpooled_h\text{bin\_size}_w = \frac{\text{roi\_w}}{\text{pooled\_w}}, \quad \text{bin\_size}_h = \frac{\text{roi\_h}}{\text{pooled\_h}}

binx1=clamp(⌊pw⋅bin_sizew⌋+roi_startw,0,W)\text{bin}_{x1} = \text{clamp}(\lfloor pw \cdot \text{bin\_size}_w \rfloor + \text{roi\_start}_w, 0, W)

binx2=clamp(⌈(pw+1)⋅bin_sizew⌉+roi_startw,0,W)\text{bin}_{x2} = \text{clamp}(\lceil (pw+1) \cdot \text{bin\_size}_w \rceil + \text{roi\_start}_w, 0, W)

(The same applies to the y direction, clamped to [0, H])

floor/ceil is based on the relative offset (excluding roi_start), and then the integer roi_start is added.

4. Max pooling (empty bins output 0):

y[n,c,ph,pw]={0if binx2≤binx1 or biny2≤biny1max⁡h,w∈[biny1,biny2)×[binx1,binx2)x[batch_idx,c,h,w]otherwisey[n, c, ph, pw] = \begin{cases} 0 & \text{if } \text{bin}_{x2} \le \text{bin}_{x1} \text{ or } \text{bin}_{y2} \le \text{bin}_{y1} \\ \max_{h, w \in [\text{bin}_{y1}, \text{bin}_{y2}) \times [\text{bin}_{x1}, \text{bin}_{x2})} x[\text{batch\_idx}, c, h, w] & \text{otherwise} \end{cases}

The +1 offset is on the ROI size (roi_width = roi_end - roi_start + 1), which is the common standard of ROI Pooling (originating from Fast R-CNN)

Parameters

Parameter name Input/Output/Attribute Description Data type Data format
x Input Feature map, with shape [N,C,H,W]; it is x in the formulas. FLOAT, FLOAT16 ND
rois Input ROI boxes, with shape [K,5]; each row is [batch_idx,x1,y1,x2,y2]. FLOAT, FLOAT16 ND
roi_actual_num Input Actual number of ROIs of each batch; optional input that does not participate in the computation in the current version. INT32 ND
y Output Pooling result, with shape [K,C,pooled_h,pooled_w]; it is y in the formulas. FLOAT, FLOAT16 ND
pooled_h Attribute Height of the pooling output. Int -
pooled_w Attribute Width of the pooling output. Int -
spatial_scale_h Attribute Scaling factor in the height direction. Float -
spatial_scale_w Attribute Scaling factor in the width direction. Float -

Constraints

  • The input x supports only 4 dimensions [N,C,H,W].
  • The input rois supports only 2 dimensions [K,5]; each row is [batch_idx,x1,y1,x2,y2].
  • roi_actual_num is an optional input and does not participate in the computation in the current version.
  • pooled_h and pooled_w must be greater than 0.
  • spatial_scale_h and spatial_scale_w must be greater than 0.
  • The dtypes of x and rois must be consistent.
  • The shape of the output y is [K,C,pooled_h,pooled_w].
  • ROI coordinates are rounded to int with roundf, and the +1 offset is on the ROI size (roi_width = roi_end - roi_start + 1).
  • Malformed ROIs are forced to be non-empty, and roi_size takes max(size,1) (int arithmetic).
  • Empty bins output 0.

Invocation Description

Invocation method Invocation sample Description
Graph-mode invocation test_geir_roi_pooling See operator invocation to complete operator compilation and verification.