Skip to content

Commit d9470e7

Browse files
AlexBodnerclaude
AlexBodner
and
claude
committed
docs(reid): document fusion choice and the proximity gate
Documents reid_fusion and its parameters, and compares both fusion rules against a geometry-only baseline on SoccerNet, DanceTrack and MOT17 on the published tuned BoT-SORT geometry. The geometry-only column is the one that decides whether to enable appearance at all, so it belongs in the table. Adds a section on reid_proximity_threshold. Its 0.5 default restricts appearance to pairs already overlapping at IoU >= 0.5, and 0.99 still requires IoU >= 0.01, which a target returning from outside the frame does not meet: its prediction has been extrapolated out of frame and overlaps the returning detection by exactly zero. Only 1.0 disables the gate. On SoccerNet test that is worth +1.23 HOTA and a third fewer ID switches on the published tuned geometry, and it is what lets a player who walked out of frame come back with the same id. The reference Deep OC-SORT recovers that case with a geometry-only pass against the last observed box; BoT-SORT has no such stage, so appearance at an open gate is the only route. The same change hurts DanceTrack, whose targets stay in frame, so this is documented as domain guidance rather than a new default. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
1 parent 199b8eb commit d9470e7

2 files changed

Lines changed: 73 additions & 6 deletions

File tree

‎CHANGELOG.md‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -8,6 +8,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
88

99
### 🚀 Added
1010

11+
- **`BoTSORTTracker(reid_fusion="adaptive")`** — Deep OC-SORT's adaptive appearance weighting as an alternative to BoT-SORT's `min(d_iou, d_app)` fusion: adds a weighted appearance term whose weight grows when the best match stands clear of the runner-up. New `reid_appearance_weight` and `reid_adaptive_weight_cap` parameters; default fusion is unchanged. `reid_appearance_threshold` has no effect under adaptive fusion. `fuse_adaptive_reid_association` lives in `trackers.core.reid.fusion` beside the BoT-SORT rule so other trackers can reuse it ([#571](https://github.com/roboflow/trackers/pull/571)).
1112
- **`trackers benchmark mcbyte` subcommand** — runs McByte over complete MOT17, DanceTrack, SportsMOT, or SoccerNet-tracking benchmark test sets and writes MOTChallenge-format results, with flags for dataset selection (`--dataset`), per-dataset detection/image roots (`--dataset_roots`), device, output location, CMC method/downscale, isolated mask matching, skip-existing, and partial-result retention ([#541](https://github.com/roboflow/trackers/pull/541), [#543](https://github.com/roboflow/trackers/pull/543)).
1213
- **`trackers inspect` command group** — visual validation commands for the mask stack and the tracker that uses it: `inspect sam` (box-prompted mask generation), `inspect cutie` (mask propagation), `inspect mask-manager` (mask lifecycle over a frame range, driven either from command-line boxes or from a MOT ground-truth file), and `inspect mcbyte` (locked-IoU baseline against full mask-conditioned McByte). Each writes annotated per-frame images into a timestamped run directory ([#543](https://github.com/roboflow/trackers/pull/543)).
1314

‎docs/guides/reid.md‎

Lines changed: 72 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -47,12 +47,15 @@ For the model catalog and fine-tuning, see the [`reid` training guide](https://g
4747

4848
## Key Parameters
4949

50-
| Parameter | Purpose | Tuning guidance |
51-
| --------------------------- | ------------------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------- |
52-
| `reid_model` | Appearance encoder queried during association. | Leave unset for IoU and CMC only. Pick a checkpoint trained on your object domain where possible. |
53-
| `reid_ema_alpha` | EMA momentum for a track's appearance feature. | Default 0.9. Higher keeps a stable long-term identity; lower adapts faster to appearance change but drifts more. |
54-
| `reid_appearance_threshold` | Maximum appearance distance `d_app` for appearance to lower a pair's matching cost. | BoT-SORT paper default 0.25. Calibrate per encoder and domain, see below. |
55-
| `reid_proximity_threshold` | IoU gate applied before appearance (`IoU ≥ 1 - reid_proximity_threshold`), from true IoU even with GIoU/DIoU/CIoU. | Default 0.5. Lower restricts how far apart a pair may be before appearance stops contributing. |
50+
| Parameter | Purpose | Tuning guidance |
51+
| --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------- |
52+
| `reid_model` | Appearance encoder queried during association. | Leave unset for IoU and CMC only. Pick a checkpoint trained on your object domain where possible. |
53+
| `reid_ema_alpha` | EMA momentum for a track's appearance feature. | Default 0.9. Higher keeps a stable long-term identity; lower adapts faster to appearance change but drifts more. |
54+
| `reid_appearance_threshold` | Maximum appearance distance `d_app` for appearance to lower a pair's matching cost. | BoT-SORT paper default 0.25. Calibrate per encoder and domain, see below. |
55+
| `reid_proximity_threshold` | IoU gate applied before appearance (`IoU ≥ 1 - reid_proximity_threshold`), from true IoU even with GIoU/DIoU/CIoU. | Default 0.5. Raise to 1.0 where targets leave the frame and return, see [below](#choosing-a-proximity-threshold). |
56+
| `reid_fusion` | How appearance combines with geometry: `"botsort"` takes the minimum of the two costs, `"adaptive"` adds a weighted appearance term. | Default `"botsort"`. See [choosing a fusion method](#choosing-a-fusion-method) before switching. |
57+
| `reid_appearance_weight` | Base appearance weight when `reid_fusion="adaptive"`. Ignored otherwise. | Default 0.75. Raise where geometry is unreliable, see below. |
58+
| `reid_adaptive_weight_cap` | Ceiling on the adaptive bonus when `reid_fusion="adaptive"`. Ignored otherwise. | Default 0.5. Raise together with `reid_appearance_weight`. |
5659

5760
---
5861

@@ -131,6 +134,69 @@ The cross-domain encoder fails differently. On SoccerNet the different-ID rate a
131134

132135
---
133136

137+
## Choosing a proximity threshold
138+
139+
`reid_proximity_threshold` decides which track-detection pairs appearance is allowed to score. A pair is dropped before appearance is consulted whenever `1 - IoU` exceeds the threshold, so the 0.5 default limits appearance to pairs that already overlap at `IoU >= 0.5`, and 0.99 still requires `IoU >= 0.01`. Only 1.0 disables the gate.
140+
141+
**Why it matters.** Lost tracks are represented by their Kalman prediction. After an occlusion the prediction has drifted and overlaps the re-emerging detection weakly; after a target leaves the frame the prediction has been extrapolated out of it and the overlap with the returning detection is exactly zero. The default excludes both cases, 0.99 admits the first, and only 1.0 admits the second. The reference Deep OC-SORT tracker recovers the second case with a geometry-only pass against the last observed box (OCR); BoT-SORT has no such stage, so appearance at an open gate is the only route.
142+
143+
On SoccerNet test (oracle detections, `osnet_x1_0` fine-tuned on SoccerNet train, library-default geometry):
144+
145+
| `reid_proximity_threshold` | appearance consulted when | HOTA | ID switches |
146+
| :------------------------- | :------------------------ | :---: | :---------: |
147+
| none (no ReID) | | 84.56 | 2939 |
148+
| 0.5 (default) | `IoU >= 0.50` | 84.52 | 3012 |
149+
| 0.8 | `IoU >= 0.20` | 85.02 | 3268 |
150+
| 0.95 | `IoU >= 0.05` | 85.69 | 2565 |
151+
| 0.99 | `IoU >= 0.01` | 85.82 | 2431 |
152+
153+
On the [published tuned BoT-SORT geometry](../evaluations/results.md#soccernet-tracking) for SoccerNet, 0.99 scores 85.87 HOTA with 2433 switches and 1.0 scores **86.23** with **1692**, against 85.00 and 2523 without appearance. The identity gain is almost entirely the last step: 1.0 is what lets a player who walked out of frame come back with the same id.
154+
155+
Half-opening is the worst setting for identity: 0.8 has more ID switches than either the closed gate or no appearance at all, because it admits enough distant candidates to create false matches without admitting the ones that enable recoveries. Either keep the gate or disable it.
156+
157+
The right value depends on the footage. DanceTrack val targets stay in frame, and there the best adaptive configuration falls from 57.26 HOTA at 0.5 to 55.88 at 0.9 and 54.89 at 1.0 (reported weights). Use 1.0 where identity is lost to camera motion or targets leaving the frame; keep 0.5 where targets stay in view.
158+
159+
## Choosing a fusion method
160+
161+
`reid_fusion` selects how the appearance score reaches the cost matrix.
162+
163+
`"botsort"` (default) takes `min(d_iou, d_app)`. Appearance competes with geometry and wins only when it is strictly cheaper, so a pair is accepted when either cue is confident. Both gates are hard: a pair either clears `reid_appearance_threshold` and `reid_proximity_threshold` or contributes nothing.
164+
165+
`"adaptive"` adds a weighted appearance term to the geometric similarity, with the weight growing when the best appearance match stands clear of the runner-up and falling back to `reid_appearance_weight` when the top candidates are hard to tell apart. This is the weighting from Deep OC-SORT, ported on its own onto BoT-SORT; the velocity term, the confidence-gated feature update and the OCR pass are not included. `reid_appearance_threshold` has no effect under this method.
166+
167+
Two consequences are worth knowing before switching:
168+
169+
- **The similarity range changes.** `"botsort"` returns values in `[0, 1]`; `"adaptive"` returns `[0, 1 + reid_appearance_weight + reid_adaptive_weight_cap]`, which is `[0, 2.25]` at the defaults. `minimum_iou_threshold_first_assoc` is applied to that fused value, so a threshold tuned for one method is a different gate under the other. Retune it when you switch, or the comparison measures the gate rather than the fusion.
170+
- **The weights are domain-dependent.** Deep OC-SORT reports `reid_appearance_weight=0.75` with `reid_adaptive_weight_cap=0.5` for MOT17 and MOT20, and `1.25` with `1.0` for DanceTrack, where dancers occlude constantly and geometry carries less. The defaults here follow the MOT17/MOT20 pair.
171+
172+
### Measured comparison
173+
174+
HOTA, best configuration found for each method. Detections, CMC, encoder and geometry are shared within a row; only appearance handling differs. SoccerNet and MOT17 use the published tuned BoT-SORT geometry from the [tracker comparison](../evaluations/results.md); DanceTrack's tuned set coincides with the library defaults.
175+
176+
| Dataset | Encoder | no ReID | `botsort` | `adaptive` | `adaptive` parameters |
177+
| -------------- | --------------------------------- | --------- | --------- | ---------- | ---------------------------------------------------------- |
178+
| SoccerNet test | `osnet_x1_0` fine-tuned SoccerNet | 85.00 | 85.46 | **86.23** | `reid_proximity_threshold=1.0` |
179+
| DanceTrack val | `osnet_x1_0_msmt17_combineall` | 53.89 | 56.11 | **57.39** | `reid_appearance_weight=2.4`, `reid_adaptive_weight_cap=0` |
180+
| MOT17 val-half | `osnet_x1_0` fine-tuned MOT17 | **69.05** | 69.00 | 68.72 | defaults |
181+
182+
SoccerNet's `botsort` figure uses `reid_appearance_threshold=0.075` and `reid_proximity_threshold=0.99`. MOT17 moves by less than 0.35 HOTA in any direction on 7 sequences, inside single-sequence variance.
183+
184+
**`adaptive` is the better fusion rule where appearance helps.** It leads on SoccerNet and DanceTrack and cuts ID switches on both (SoccerNet 2523 to 1692, DanceTrack 1905 to 1683), where `botsort` raises them on DanceTrack (2276) while still gaining HOTA. The advantage comes from adding appearance instead of taking a minimum: with `reid_adaptive_weight_cap=0`, which disables the adaptive bonus, HOTA changes by 0.12 on SoccerNet and 0.13 on DanceTrack. Leave the cap at its default and tune `reid_appearance_weight` and `reid_proximity_threshold` instead.
185+
186+
### Choosing the adaptive weight
187+
188+
Deep OC-SORT reports `reid_appearance_weight=0.75` for MOT17 and MOT20 and `1.25` for DanceTrack. On SoccerNet test and MOT17 val-half a sweep from 0.75 to 2.5 moves HOTA by less than 0.15, so the default transfers. DanceTrack val needs more than the reported value:
189+
190+
| `reid_appearance_weight` | HOTA | ID switches |
191+
| ------------------------ | --------- | ----------- |
192+
| 1.25 (reported) | 55.97 | 1736 |
193+
| 1.6 | 57.23 | 1696 |
194+
| 2.4 | **57.26** | **1686** |
195+
196+
The step sits between 1.25 and 1.6 and everything above it is a plateau. Where appearance barely separates identities the adaptive bonus collapses and only the base weight does any work, which is DanceTrack's situation and why the weight matters there.
197+
198+
`"botsort"` remains the default because it is the more conservative of the two: bounded output, a hard appearance gate, and no threshold changes when you enable it.
199+
134200
## BoT-SORT with and without ReID
135201

136202
### MOT17 test

0 commit comments

Comments
 (0)