🧩 How the detector works

One pass through the network turns an image into a fixed list of boxes. This page follows that pass in five stages and says which source file owns each one. The numbers on the NMS comparison page and in the reproduction report come from exactly this path.

Each stage lights up in turn; with reduced motion enabled the sketch stays fully lit. It shows the data path, not measured output.

🔎 The five stages

1. Image. A letterboxed square, 640 px for detection, segmentation and keypoints, 1024 px for oriented detection tiles.
2. Backbone and neck. One trunk, shared by all four tasks, turns the image into feature grids at strides 8, 16 and 32. A stride-8 cell sees a small patch and decides about small objects; a stride-32 cell sees a large patch and decides about large ones.
3. Two heads on the same grids. Every grid cell is an anchor point. Each head predicts, for each anchor, class scores and four distances (left, top, right, bottom) from the anchor to the box edges. There are no distribution bins to decode: the distances are the box (models/heads/detect.py).
4. Top-k. The deployed head's class scores are ranked and the best 300 anchor predictions are kept (decode/topk_e2e.py), regardless of how many real objects are present. The output shape is always (batch, 300, 6): [x1, y1, x2, y2, score, class]. Zero padding is needed only when fewer than 300 anchors exist; callers can filter predictions by confidence.
5. Boxes. That tensor is the answer. Nothing after it merges or filters boxes.

🪞 Why two heads

The two heads read the same features and differ in what they are taught during training.

	one-to-many head	one-to-one head
Training target	several anchors per object count as positives	exactly one anchor per object counts as positive
Effect	dense, strong gradient signal for the shared trunk	learns to fire once per object
At inference	optional: feeds the NMS path	deployed: feeds top-k

Training blends the two losses and shifts weight from the dense head to the sparse one as the run progresses (Progressive Loss, losses/progressive.py: alpha ramps from 0.8 to 0.1 over the epoch budget). The dense head is why the trunk learns quickly; the sparse head is why inference needs no cleanup.

🧭 Where to look in the code

Stage	Source
heads, ltrb boxes	models/heads/detect.py
assignment (which anchors are positive)	assign/tal.py, assign/stal.py
NMS-free decode	decode/topk_e2e.py
classic decode, for comparison	decode/nms_path.py

Every mechanism above is traced to a numbered section of the source paper in the reproduction report's What was reproduced table. What the project had to assume where the paper is silent is in the assumption register.

Next: what removing NMS costs.