Non-maximum suppression in PyTorch - Model Pipeline Trace

Model Pipeline - Non-maximum suppression

Non-maximum suppression (NMS) is a technique used in object detection to remove overlapping bounding boxes. It keeps only the most confident box for each object, helping the model make clear decisions.

Data Flow - 3 Stages

1Input bounding boxes and scores

1000 boxes x 4 coordinates + 1000 scores→Receive predicted bounding boxes and their confidence scores from the detection model→1000 boxes x 4 coordinates + 1000 scores

Boxes: [[10, 20, 50, 60], [12, 22, 48, 58], ...], Scores: [0.9, 0.85, ...]

↓

2Sort boxes by confidence score

1000 boxes + scores→Sort all boxes from highest to lowest confidence score→1000 boxes + scores sorted

Sorted scores: [0.95, 0.9, 0.85, ...], corresponding boxes reordered

↓

3Apply Non-maximum suppression

1000 boxes + scores sorted→Iteratively select the highest score box and remove boxes that overlap too much (IoU > threshold)→Filtered boxes + scores (less than or equal to 1000)

Kept boxes: [[10, 20, 50, 60], [100, 120, 150, 160], ...], Scores: [0.95, 0.88, ...]

Training Trace - Epoch by Epoch

Loss
1.2 |****
0.9 |***
0.7 |**
0.5 |*
0.4 |

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Initial training with high loss and low accuracy
2	0.9	0.60	Loss decreased, accuracy improved as model learns
3	0.7	0.72	Better bounding box predictions, NMS helps reduce duplicates
4	0.5	0.80	Model converging, NMS effectively filters overlapping boxes
5	0.4	0.85	Final epoch with good balance of precision and recall

Prediction Trace - 3 Layers

Layer 1: Input bounding boxes and scores

Layer 2: Sort boxes by scores

Layer 3: Non-maximum suppression

Model Quiz - 3 Questions

What is the main purpose of Non-maximum suppression in object detection?

A To randomly select boxes for training

B To increase the number of predicted boxes

C To remove overlapping boxes and keep the most confident ones

D To sort boxes by size

Key Insight

Non-maximum suppression helps object detection models by removing duplicate overlapping boxes, keeping only the most confident predictions. This improves the clarity and quality of final detections.