docs(reid): what ReID costs in frames per second

[AlexBodner]

Adds a Speed section to the ReID guide, comparing BoT-SORT without ReID, with each dataset's fine-tuned osnet_x1_0, and with the published fastreid_mot17_sbs50. Stacked on #571.

The benchmark notebook contains the benchmark code and saved outputs used for the results below. Rerunning it requires access to the benchmark inputs and fine-tuned encoders.

Dataset	Split	BoT-SORT (FPS)	+ fine-tuned OSNet (FPS)	+ FastReID (FPS)
MOT17	val-half	31.1	10.7	6.1
SportsMOT	val	38.4	16.3	10.3
DanceTrack	val	33.8	15.4	10.9
SoccerNet	train	32.6	13.2	6.8

Adding the fine-tuned OSNet made tracking 2.2 to 2.9 times slower than BoT-SORT alone. Adding FastReID made it 3.1 to 5.1 times slower.

How it was measured

• Timed tracker.update only. Frame decoding and detection parsing were outside the timed call; encoder loading happened at construction and was not timed.
• BoT-SORT sent only high-confidence detections to the encoder. All three runs for a sequence used the same provided detections and library defaults, with camera motion compensation on. SoccerNet used ground-truth boxes, as in the Results tables.
• Measured every sequence of each tuning split: 134 sequences and 97,880 frames. The notebook saved the first three datasets in its main run and SoccerNet in a separate completed cell.
• The notebook setup output reports a Google Colab Tesla T4, an Intel Xeon @ 2.00GHz with 2 vCPUs, and PyTorch 2.11.0+cu130. These values replace an earlier T4 run using PyTorch 2.11.0+cu128; results from the two runs are not mixed in the table.

[Copilot]

🟡 Changes recommended

The methodology incorrectly states that ReID encodes every detection rather than only high-confidence detections.

1 open finding

• Clarify that ReID embeddings use only high-confidence boxes · New

What changed in this PR

Adds measured ReID throughput costs to the BoT-SORT ReID guide.

Changes:

• Documents benchmark methodology and hardware.
• Compares FPS across four datasets and two encoders.
• Summarizes observed throughput reductions.

File	Description
docs/​guides/​reid.md	Adds the ReID speed benchmark section and results.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[tstanczyk95]

@AlexBodner It's a promising table, but it needs clarification. See my remarks above.