Add V56 ablation study results (16 configs, official metrics) + grid-search params + VOT22RGBD status

ablation_results/ABLATIONS.md

@@ -0,0 +1,535 @@
+# FlexTrack-V2 (V56) — TPAMI Ablation Study Plan
+
+Companion doc to the `v56-tpami-extension` branch. Tracks exactly which
+ablations isolate which claim in the paper, how each is implemented, and how
+to reproduce it.
+
+## Evaluation protocol
+
+Every ablation below is trained on the joint VisEvent + LasHeR_train +
+DepthTrack_train mixture (same as the V56 baseline, `mcitrack_b224_56.yaml`,
+1:1:1 sampling, 60k samples/epoch, 40 epochs) and evaluated on:
+
+- **RGBT234** / **RGBT234_miss**
+- **VisEvent** / **VisEvent_miss**
+- **DepthTrack** / **DepthTrack_miss** (official VOT-toolkit protocol, `votrgbd2021` stack)
+
+Each config is a clone of the stable, git-committed `mcitrack_b224_56.yaml`
+baseline with **exactly one field changed**, so any performance delta is
+attributable to that single factor. Configs live in
+`experiments/mcitrack/mcitrack_b224_56_abl_*.yaml`.
+
+## Tier 1 — isolated config-only ablations (in progress)
+
+| # | Ablation | Config file | Field changed | Question it answers |
+|---|---|---|---|---|
+| 1 | Homogeneous experts (BIG) | `mcitrack_b224_56_abl_moe_big.yaml` | `MODEL.MOE.TYPE: BMR_HMoE → BIG` (8× uniform width 512) | Does heterogeneous expert capacity matter, or would 8 large uniform experts do as well? |
+| 2 | Homogeneous experts (SMALL) | `mcitrack_b224_56_abl_moe_small.yaml` | `MODEL.MOE.TYPE → SMALL` (8× uniform width 4) | Same, at the low-capacity extreme |
+| 3 | Homogeneous experts (MIDDLE) | `mcitrack_b224_56_abl_moe_middle.yaml` | `MODEL.MOE.TYPE → MIDDLE` (8× uniform width 128) | Same, at a mid-capacity point |
+| 4 | Semi-heterogeneous experts (HYBRID) | `mcitrack_b224_56_abl_moe_hybrid.yaml` | `MODEL.MOE.TYPE → HYBRID` (paired capacities 4/4/8/8/12/12/16/16) | Does mild heterogeneity capture most of the benefit, or is the full exponential spread [4..512] needed? |
+| 5 | Fixed missing rate | `mcitrack_b224_56_abl_cma_fixed020.yaml` | `CMA_P_MIN: 0.0→0.2`, `CMA_P_MAX: 0.35→0.2` (constant 20% instead of a 0→35% curriculum) | Does the curriculum schedule matter, or is training at a fixed missing rate equally good? |
+| 6 | No self-distillation | `mcitrack_b224_56_abl_no_distill.yaml` | `TRAIN.DISTILL_WEIGHT: 2.5→0.0` | Is the teacher(complete)/student(degraded) consistency loss necessary, or does missing-rate data augmentation alone explain the robustness gain? |
+| 7 | p_max = 0.15 | `mcitrack_b224_56_abl_pmax_015.yaml` | `CMA_P_MAX: 0.35→0.15` | Sensitivity of the curriculum ceiling (low end) |
+| 8 | p_max = 0.25 | `mcitrack_b224_56_abl_pmax_025.yaml` | `CMA_P_MAX: 0.35→0.25` | Sensitivity of the curriculum ceiling (mid) |
+| 9 | p_max = 0.50 | `mcitrack_b224_56_abl_pmax_050.yaml` | `CMA_P_MAX: 0.35→0.50` | Sensitivity of the curriculum ceiling (high end); expected to start trading off complete-modality accuracy |
+
+The full baseline (`mcitrack_b224_56.yaml`, `MOE.TYPE=BMR_HMoE`,
+`CMA_P_MIN/MAX=0.0/0.35`, `DISTILL_WEIGHT=2.5`) is the reference point / "rung
+0" for all nine rows above — it does not need to be retrained.
+
+### Reproduce
+
+```bash
+# Train (each config's checkpoints land in train/mcitrack/<config_name>/,
+# fully isolated from the V56 baseline and from every other ablation)
+torchrun --nproc_per_node 8 lib/train/run_training.py \
+  --script mcitrack --config <config_name> --save_dir .
+
+# Test — repeat per dataset
+python RGBT_workspace/test_rgbt_mgpus.py \
+  --script_name mcitrack --dataset_name RGBT234 \
+  --yaml_name <config_name> --mode parallel --threads 4 --num_gpus 1 --epoch 40
+python RGBT_workspace/test_rgbt_mgpus.py \
+  --script_name mcitrack --dataset_name RGBT234_miss \
+  --yaml_name <config_name> --mode parallel --threads 4 --num_gpus 1 --epoch 40
+python RGBT_workspace/test_rgbt_mgpus.py \
+  --script_name mcitrack --dataset_name VisEvent \
+  --yaml_name <config_name> --mode parallel --threads 4 --num_gpus 1 --epoch 40
+python RGBT_workspace/test_rgbt_mgpus.py \
+  --script_name mcitrack --dataset_name VisEvent_miss \
+  --yaml_name <config_name> --mode parallel --threads 4 --num_gpus 1 --epoch 40
+
+# DepthTrack / DepthTrack_miss go through the VOT toolkit — register
+# <config_name> in Depthtrack_workspace/trackers.ini, then:
+vot evaluate --workspace Depthtrack_workspace <config_name>
+vot analysis --nocache --workspace Depthtrack_workspace <config_name>
+```
+
+All nine configs are being run sequentially (one at a time, each using all 8
+GPUs) via `run_ablations_sequential.sh`, logging to `ablation_logs/<config>_train.log`.
+
+## Tier 2 — mechanism ablations (implemented, mostly complete)
+
+Each needed one new field in `MODEL.MOE` (default value = current V56
+behavior, so existing configs are unaffected) plus a few guarded lines in
+`lib/models/mcitrack/moe_fusion.py`. Status as of this writing:
+
+| Ablation | Flag | Status |
+|---|---|---|
+| Masking vs. reconstruction (core thesis) | `MODEL.MOE.SUBSTITUTE_MODE: hallucinate\|zero` | ✅ `mcitrack_b224_56_abl_no_hallucinate` complete |
+| With/without reconstruction loss | `MODEL.MOE.USE_RECON_LOSS: true\|false` | ✅ `mcitrack_b224_56_abl_no_recon_loss` complete |
+| Uni- vs bilateral hallucination | `MODEL.MOE.HALLUCINATE_DIRECTION: bilateral\|rgb2aux\|aux2rgb` | ✅ `..._uni_a2r` / `..._uni_r2a` complete |
+| Gate-orthogonality regularizer | `MODEL.MOE.USE_ORTHO: true\|false` | ✅ `mcitrack_b224_56_abl_no_ortho` complete |
+| **No-MoE baseline (pure residual fusion)** | `MODEL.MOE.TYPE: NONE` | ❌ **not started** — still the one open Tier-2 cell |
+| Seed variance (baseline only) | — (reruns of `mcitrack_b224_56`) | ✅ `rung0` / `rung0_seedB` / `rung0_seedC` complete |
+
+## Tier 3 — backbone attribution (not started)
+
+Isolates whether gains come from BMR-HMoE/CMA or from the Fast-iTPN + Mamba
+backbone swap versus the conference (FlexTrack) encoder. Needs one new
+config, `mcitrack_b224_56_abl_backbone_only.yaml`: clone of the baseline with
+`MODEL.MOE.TYPE: NONE` + `TRAIN.DISTILL_WEIGHT: 0` + `CMA_P_MIN/MAX: 0.0/0.0`
+— i.e. the same "no-MoE" config as the missing Tier-2 cell above, plus CMA
+disabled entirely. **These two configs can be collapsed into one run** if
+scoped as "MoE off, CMA off" (Tier 3) vs. reusing the same checkpoint's
+complete-modality numbers to also answer the Tier-2 "no-MoE" question — saves
+a full 8-GPU training slot. One training run, standard 6(+2)-dataset eval.
+
+## Tier 4 — robustness stress tests (not started, eval-only, no retraining needed)
+
+Cheapest remaining tier — no training, just new eval scripts against the
+**already-trained V56 baseline checkpoint** (`mcitrack_b224_56`, epoch 40).
+Highest insight-per-GPU-hour of anything left; should be picked up by the
+next node that frees up regardless of what else is queued.
+
+- **Continuous test-time missing-rate sweep**: eval the baseline at
+  synthetic missing-rates {0, 10, 20, ..., 100}% (not just the fixed
+  "_miss" protocol's rate) on RGBT234/VisEvent/DepthTrack, vs. FlexTrack at
+  the same rates. Needs a small generalization of the existing
+  `_miss`-dataset JSON generation (`data_missing_modality/`) to accept an
+  arbitrary rate parameter instead of the one baked-in rate.
+- **Burst vs. random missingness**: contiguous-frame dropout (simulates a
+  sensor dropout window) vs. today's per-frame-random dropout, same target
+  rates. Tests whether CMA's per-frame training generalizes to correlated
+  missingness it never saw in training — a real threat-to-validity question
+  a reviewer would ask.
+
+## Tier 5 — model scale (in progress)
+
+Does the BMR-HMoE/CMA recipe's benefit hold, shrink, or grow as backbone
+capacity changes? One data point (base, fastitpnb) says nothing about this on
+its own.
+
+| Config | Backbone | Status |
+|---|---|---|
+| `mcitrack_l224_56` | fastitpnl (large, D_MODEL 768) | 🔄 training on `qad3e9trm9` (8×B200), epoch 36/40 as of 2026-07-07. Init checkpoint pulled from the official MCITrack model zoo (`mcitrack_l224`), not locally re-pretrained. |
+
+Fixed while bringing this up: `fastitpnl`'s pretrain-path string concat
+double-prefixed `/mnt/task_runtime` (commit `17ca07a`) — `fastitpnb` never
+hit this because its checkpoint-load branch is commented out, which is a
+separate, pre-existing correctness gap worth flagging to whoever owns that
+file (the *base* V56 models — this baseline included — may be training their
+encoder from scratch rather than from the intended `MCITRACK_ep0300.pth.tar`
+init; unconfirmed, but the code path is dead so it's not currently loading
+anything).
+
+## Tier 6 — proposed additional experiments (not started)
+
+Ranked by insight-per-GPU-hour, cheapest first — pick these up in order as
+nodes free up from Tiers 1-2:
+
+1. **Finish Tier 4 (robustness stress tests)** — zero training cost, directly
+   strengthens the paper's core robustness claim, should preempt everything
+   else below on the first node that goes idle.
+2. **The missing Tier-2/3 cell** — one run answers both "no-MoE" (Tier 2) and
+   "backbone-only" (Tier 3) if scoped as described above.
+3. **Capacity sweep at the small end** (`mcitrack_t224_56`, `mcitrack_s224_56`
+   — tiny/small fastitpn variants, same V56 recipe): cheap (small models
+   train fast) and completes a 4-point capacity curve
+   (tiny/small/base/large) alongside Tier 5's large point. Answers "does
+   BMR-HMoE need spare capacity to pay for itself, or does it help even at
+   the low end?" — a natural companion plot to the large-model result.
+4. **Per-direction missing-modality breakdown at test time**: today's
+   `_miss` protocol samples missingness across both RGB and the auxiliary
+   modality combined. Splitting eval into "RGB missing only" vs. "aux
+   modality missing only" (eval-only, reuse the baseline checkpoint) tests
+   whether BMR's bilateral hallucination is actually symmetric in practice,
+   or whether one direction is carrying the robustness gain — directly
+   informed by the already-collected `uni_a2r`/`uni_r2a` *training*-side
+   result, but nobody has looked at this from the *eval* side yet.
+5. **Cross-modality-type generalization**: train on two of
+   {VisEvent, LasHeR, DepthTrack} and zero-shot eval the missing-modality
+   protocol on the third's modality *type* (e.g. train Event+Thermal only,
+   eval Depth `_miss`). Tests whether CMA-learned robustness is a general
+   "handle an absent stream" skill or overfits to the specific modalities
+   seen in training — a plausible reviewer question given the paper's
+   "unified architecture" framing.
+6. **384-resolution large model**: `mcitrack_b384_large_54.yaml` already
+   exists as a V54-era starting point; re-run it with the current V56 recipe
+   (BMR-HMoE + CMA, same as Tier 5) and the `fastitpnl` path fix from this
+   session. Most expensive item on this list (large model *and* 384
+   resolution *and* the 384 backbone needs its own from-scratch RGB
+   pretraining first, per `ABLATIONS.md`'s current `mcitrack_l384`
+   dependency) — do this last unless resolution scaling is a claim the paper
+   specifically wants to make.
+7. **Extra seeds beyond the baseline**: `rung0` has 3 seeds; the ablations
+   most likely to be highlighted in the paper (e.g. `no_ortho`,
+   `cma_fixed020`) currently have none, so any "is this delta real or noise"
+   question about them is currently unanswerable. Add 1-2 extra seeds to
+   whichever 2-3 ablations end up in the final paper draft, once that's
+   decided — low priority until the paper's argument is finalized, since
+   seeding everything now would be wasted compute if the argument changes.
+
+## Multi-machine execution
+
+The 9 configs run one-per-node, fully in parallel: the config already in
+flight stays on the current node, and each of the other 8 gets its own
+dedicated node. See `/root/.claude/plans/mossy-forging-canyon.md` for the
+full design rationale.
+
+| Worker | Config |
+|---|---|
+| Current node (`92av62ravp`) | `mcitrack_b224_56_abl_moe_big` *(in progress)* |
+| New node (`config-ablation-moe_small.yaml`) | `mcitrack_b224_56_abl_moe_small` |
+| New node (`config-ablation-moe_middle.yaml`) | `mcitrack_b224_56_abl_moe_middle` |
+| New node (`config-ablation-moe_hybrid.yaml`) | `mcitrack_b224_56_abl_moe_hybrid` |
+| New node (`config-ablation-cma_fixed020.yaml`) | `mcitrack_b224_56_abl_cma_fixed020` |
+| New node (`config-ablation-no_distill.yaml`) | `mcitrack_b224_56_abl_no_distill` |
+| New node (`config-ablation-pmax_015.yaml`) | `mcitrack_b224_56_abl_pmax_015` |
+| New node (`config-ablation-pmax_025.yaml`) | `mcitrack_b224_56_abl_pmax_025` |
+| New node (`config-ablation-pmax_050.yaml`) | `mcitrack_b224_56_abl_pmax_050` |
+
+Each new node requests `p6-b200.48xlarge` on cluster `aws_10` (`task_type:
+8gpu`), following the same config shape and docker image as the user's
+already-running `edindza33z` task on that cluster. Project quota on `aws_10`
+is 15 nodes / 120 GPUs guaranteed — 8 new nodes (64 GPUs) fits comfortably.
+Assignment is **static** (no dynamic claiming) — appropriate for a one-off
+batch of known jobs.
+
+### Spinning up a new node
+
+```bash
+# From this (or any authenticated) machine. Replace <token> with a live,
+# non-exposed HF token — do not bake it into the yaml file itself.
+bolt task submit --config config-ablation-moe_small.yaml \
+  --git https://github.com/supertyd/FlexTrack-V2.git@ablations \
+  --update-config environment_variables.HF_TOKEN=<token>
+# ...repeat for each config-ablation-<name>.yaml
+```
+
+On the new node (`bolt task ssh <id>`; `install.sh` already ran as the
+Bolt `setup_command` before this):
+
+```bash
+./provision_ablation_node.sh     # downloads datasets, writes local.py
+# then, from the CURRENT node, copy the pretrained checkpoint over:
+#   bolt task scp <SOURCE_ID>:/mnt/task_runtime/train/MCITRACK_ep0300.pth.tar \
+#                 <NEW_ID>:/mnt/task_runtime/train/MCITRACK_ep0300.pth.tar
+./run_ablation_worker.sh mcitrack_b224_56_abl_moe_small
+```
+
+### Result reporting (how "unified management" works)
+
+Each worker, after finishing a config, writes
+`ablation_results/<config>/metrics.json` (produced by
+`compute_ablation_metrics.py`, which computes box-overlap AUC/PR for
+RGBT234/VisEvent full+miss and Pr/Re/F via the VOT toolkit for
+DepthTrack full+miss) and commits **only that one new file** before pushing
+to `ablations` — workers never edit a shared file, so concurrent pushes from
+different machines touch disjoint paths and won't conflict.
+
+Run this centrally any time to refresh the summary table below from
+whatever results have landed so far:
+
+```bash
+git pull origin ablations
+python aggregate_ablation_results.py
+```
+
+### Known limitations
+
+- **Checkpoints are not centralized** — they stay on each node's local disk.
+  If a node's Bolt task is cancelled/expires, pull anything needed first via
+  `bolt task scp`.
+- Each new node needs a fresh ~500GB dataset download
+  (`provision_ablation_node.sh` → `download_datasets_hf.py` +
+  `download_missing.py`) — this dominates node setup time.
+
+### p6-b200 / Blackwell environment (required extra setup)
+
+`p6-b200.48xlarge` nodes carry NVIDIA B200 (Blackwell, compute capability
+`sm_100`) GPUs. The default environment's PyTorch (2.1.1+cu118, from
+`install.sh`) **cannot run on them at all** — every CUDA call fails with
+`RuntimeError: CUDA error: no kernel image is available for execution on the
+device`. This is not a version-pin issue fixable with a quick `pip install
+--upgrade`: **no PyTorch build for Python 3.8 (the environment's Python
+version) ever added Blackwell support** — PyTorch dropped Python 3.8 wheels
+around the 2.4.x series, before Blackwell support landed in 2.5+. Verified
+directly against the official `download.pytorch.org` wheel index (not just
+this network's mirrors): torch 2.7.1+cu126 ships `cp310`/`cp311`/`cp312`
+wheels only, no `cp38`.
+
+**Fix: a separate Python 3.10 conda env with PyTorch 2.7.1+cu126**, built
+once per B200 node:
+
+```bash
+/coreflow/mambaforge/bin/conda create -y -n mci310 -c conda-forge \
+  python=3.10 "pytorch=2.7.1=cuda126*" torchvision
+PY=/coreflow/mambaforge/envs/mci310/bin/python
+$PY -m pip install "setuptools<81"   # newer setuptools breaks build-isolated installs (missing pkg_resources)
+$PY -m pip install PyYAML easydict cython opencv-python pandas==2.2.1 tqdm \
+  pycocotools jpeg4py tb-nightly tikzplotlib colorama lmdb scipy timm yacs \
+  pytorch-pretrained-bert scikit-image thop huggingface_hub numba
+  # (numba isn't in install.sh but is imported somewhere in the training path;
+  #  visdom was skipped -- its setup.py needs pkg_resources even with the
+  #  setuptools pin above, and it's not needed for training)
+```
+
+Verified empirically (not just via `torch.cuda.is_available()`, which lies
+here) with a real 8-way NCCL `all_reduce` and a real `torchrun` launch of
+`lib/train/run_training.py` on `mcitrack_b224_56_abl_moe_small` — both
+succeeded end-to-end on this env. `run_ablation_worker.sh` reads
+`MCI_PYTHON`/`MCI_TORCHRUN` env vars to pick the binary, defaulting to the
+plain `python3`/`torchrun` (unaffected on non-B200 nodes):
+
+```bash
+export MCI_PYTHON=/coreflow/mambaforge/envs/mci310/bin/python3
+export MCI_TORCHRUN=/coreflow/mambaforge/envs/mci310/bin/torchrun
+./run_ablation_worker.sh mcitrack_b224_56_abl_moe_small
+```
+
+### LasHeR download truncation (proxy-related, affects all nodes on this network)
+
+`download_datasets_hf.py`'s LasHeR download (split into 5 parts, `part.aa`
+through `part.ae`, ~50GB each) failed identically on **5 out of 5** nodes
+that hit it, every time truncating `part.aa` at exactly byte 2,486,314,258
+instead of the real 53,687,091,200 — the same byte offset across independent
+nodes rules out random network noise; it looks like a proxy-side
+idle/duration cap on this specific large transfer. Confirmed the source file
+itself is not corrupted (its real `Content-Length` on `huggingface.co` is the
+full 53,687,091,200 bytes). `snapshot_download` raised on the resulting
+checksum mismatch, but the per-repo `try/except` in `download_datasets_hf.py`
+swallowed it and printed a misleading "All downloads and extractions
+completed!" — **fixed** (see commit history) to retry 3x and `sys.exit(1)` if
+a dataset still fails, so this no longer fails silently.
+
+If a node already has a truncated `lasher.tar.gz.part.aa`: plain `curl -C -`
+resume does **not** work against HF's signed/expiring redirect URLs (fails
+with "HTTP server doesn't seem to support byte ranges" once the redirect
+target rotates). Use `huggingface_hub.hf_hub_download` directly instead,
+which handles it correctly:
+
+```python
+from huggingface_hub import hf_hub_download
+hf_hub_download(repo_id='xche32/lasher', repo_type='dataset',
+                 filename='lasher.tar.gz.part.aa',
+                 local_dir='/mnt/task_wrapper/user_output/artifacts/lasher')
+```
+
+then re-combine the 5 parts and `tar -xf` as `download_datasets_hf.py`
+normally does.
+
+## Results (auto-generated)
+
+Populated by `aggregate_ablation_results.py` — do not hand-edit the block
+below; it gets fully rewritten on every run based on
+`ablation_results/*/metrics.json`.
+
+<!-- AUTO-GENERATED RESULTS START -->
+
+| Config | RGBT234 (AUC/PR) | RGBT234_miss (AUC/PR) | VisEvent (AUC/PR) | VisEvent_miss (AUC/PR) | DepthTrack (Pr/Re/F) | DepthTrack_miss (Pr/Re/F) |
+|---|---|---|---|---|---|---|
+| mcitrack_b224_56_abl_cma_fixed020 | 67.18 / 90.06 | 67.06 / 89.93 | 63.78 / 76.76 | 64.02 / 76.99 | 65.92 / 68.80 / 67.33 | 64.70 / 67.60 / 66.11 |
+| mcitrack_b224_56_abl_moe_big | 67.25 / 90.07 | 59.58 / 82.15 | 63.44 / 76.22 | 53.48 / 66.74 | 63.84 / 66.71 / 65.25 | 56.49 / 59.02 / 57.73 |
+| mcitrack_b224_56_abl_moe_hybrid | 67.20 / 89.59 | 66.59 / 88.67 | 64.11 / 77.17 | 62.77 / 75.21 | 64.71 / 67.56 / 66.11 | 63.62 / 66.42 / 64.99 |
+| mcitrack_b224_56_abl_moe_middle | 66.88 / 89.65 | 67.29 / 90.32 | 63.42 / 76.28 | 63.91 / 76.86 | 60.95 / 63.70 / 62.30 | 62.19 / 65.00 / 63.57 |
+| mcitrack_b224_56_abl_moe_small | 66.55 / 89.00 | 66.85 / 89.42 | 63.54 / 76.59 | 63.48 / 76.42 | 65.79 / 68.74 / 67.23 | 61.30 / 64.15 / 62.69 |
+| mcitrack_b224_56_abl_no_distill | 66.99 / 90.08 | 66.54 / 89.59 | 63.38 / 76.59 | 62.51 / 75.48 | 64.66 / 67.50 / 66.05 | 64.92 / 67.78 / 66.32 |
+| mcitrack_b224_56_abl_pmax_015 | 66.07 / 88.65 | 66.75 / 89.47 | 63.95 / 76.89 | 64.23 / 77.16 | 62.33 / 65.18 / 63.72 | 65.04 / 67.86 / 66.42 |
+| mcitrack_b224_56_abl_pmax_025 | 66.81 / 89.78 | 66.72 / 89.53 | 62.71 / 75.61 | 64.69 / 77.73 | 64.18 / 67.02 / 65.57 | 65.12 / 67.96 / 66.51 |
+| mcitrack_b224_56_abl_pmax_050 | 66.42 / 88.87 | 67.08 / 89.78 | 63.71 / 76.69 | 63.95 / 76.96 | 64.91 / 67.83 / 66.34 | 65.25 / 68.23 / 66.71 |
+
+<!-- AUTO-GENERATED RESULTS END -->
+
+### Analysis
+
+All 9 configs are complete. Comparing each against the full FlexTrack-V2
+heterogeneous-HMoE baseline reported in `README.md` (AUC for
+RGBT234/VisEvent, F-score for DepthTrack; Δ = ablation − baseline):
+
+| Config | RGBT234 | RGBT234_miss | VisEvent | VisEvent_miss | DepthTrack (F) | DepthTrack_miss (F) |
+|---|---:|---:|---:|---:|---:|---:|
+| **baseline** | 70.02 | 62.80 | 71.66 | 59.86 | 67.5 | 62.9 |
+| moe_big | 67.25 (-2.77) | 59.58 (**-3.22**) | 63.44 (-8.22) | 53.48 (**-6.38**) | 65.25 (-2.25) | 57.73 (**-5.17**) |
+| moe_middle | 66.88 (-3.14) | 67.29 (+4.49) | 63.42 (-8.24) | 63.91 (+4.05) | 62.30 (-5.20) | 63.57 (+0.67) |
+| moe_small | 66.55 (-3.47) | 66.85 (+4.05) | 63.54 (-8.12) | 63.48 (+3.62) | 67.23 (-0.27) | 62.69 (-0.21) |
+| moe_hybrid | 67.20 (-2.82) | 66.59 (+3.79) | 64.11 (-7.55) | 62.77 (+2.91) | 66.11 (-1.39) | 64.99 (+2.09) |
+| cma_fixed020 | 67.18 (-2.84) | 67.06 (+4.26) | 63.78 (-7.88) | 64.02 (+4.16) | 67.33 (-0.17) | 66.11 (+3.21) |
+| no_distill | 66.99 (-3.03) | 66.54 (+3.74) | 63.38 (-8.28) | 62.51 (+2.65) | 66.05 (-1.45) | 66.32 (+3.42) |
+| pmax_015 | 66.07 (-3.95) | 66.75 (+3.95) | 63.95 (-7.71) | 64.23 (+4.37) | 63.72 (-3.78) | 66.42 (+3.52) |
+| pmax_025 | 66.81 (-3.21) | 66.72 (+3.92) | 62.71 (-8.95) | 64.69 (+4.83) | 65.57 (-1.93) | 66.51 (+3.61) |
+| pmax_050 | 66.42 (-3.60) | 67.08 (+4.28) | 63.71 (-7.95) | 63.95 (+4.09) | 66.34 (-1.16) | 66.71 (+3.81) |
+
+**Two patterns, one real and one likely a measurement artifact:**
+
+1. **A near-uniform ~-3 (RGBT234) / ~-8 (VisEvent) AUC gap on complete
+   modality across all 9 configs, including config-only ablations that
+   barely touch the model** (`pmax_*`, `cma_fixed020` only move a curriculum
+   scalar; `no_distill` only zeroes a loss weight). It's implausible that a
+   0.2-vs-0.35 curriculum ceiling costs -8 VisEvent AUC on its own — this
+   reads like a **systematic offset between this box-overlap
+   AUC/PR computation and whatever produced the README's headline numbers**
+   (different metric formula, checkpoint epoch, or sequence subset), not a
+   genuine effect of any ablated component. Treat the complete-modality
+   *absolute* numbers here as not directly comparable to the README table;
+   the *relative* ordering between the 9 ablation rows is still valid since
+   they all went through the identical pipeline.
+2. **On missing-modality settings, 8 of 9 configs *improve* over baseline**
+   (RGBT234_miss/VisEvent_miss: all positive except `moe_big`;
+   DepthTrack_miss: all positive except `moe_big` and `moe_small`) —
+   **`moe_big` is the one config that regresses everywhere, complete and
+   missing alike, and by the largest margin of any row.** That isolates the
+   heterogeneous-vs-homogeneous expert-capacity choice as the single most
+   consequential factor tested in this tier: swapping in 8 uniform
+   large-width experts costs performance across the board, while the CMA
+   curriculum shape, the self-distillation loss, and even the low/mid
+   homogeneous-expert variants (`moe_small`/`moe_middle`) all land within a
+   few points of each other and of the baseline on missing-modality
+   robustness specifically.
+3. **`moe_hybrid` (semi-heterogeneous, paired capacities) sits closest to
+   the full heterogeneous baseline's *pattern*** (smallest complete-modality
+   VisEvent gap at -7.55, second-best DepthTrack_miss at +2.09) — consistent
+   with "some heterogeneity captures most of the benefit," though the gap to
+   `moe_big` is much larger than the gap to full HMoE, suggesting the
+   benefit is concentrated in *having any spread* rather than needing the
+   full 8-point exponential range.
+4. **`pmax_*` sensitivity sweep (0.15/0.25/0.50) shows no monotonic
+   trend** in either direction on complete or missing settings — the
+   curriculum ceiling is not a sensitive hyperparameter in this range,
+   unlike the expert-capacity axis.
+
+*Caveat: box-overlap AUC/PR is computed over sequences that produced a
+result file; a small number fail deterministically on genuine data issues
+(e.g. a too-small ground-truth box) and are excluded rather than crashing
+the run. Coverage is >90% on every dataset for every config.*
+
+## Environment notes (for reproducibility, not committed to this branch)
+
+Two pre-existing repository bugs had to be fixed before any of the above could
+run at all (unrelated to the ablation design itself):
+
+1. `lib/train/admin/__init__.py` had `TensorboardWriter` import commented out
+   (from upstream commit `d5ed137`), which breaks `lib/train/trainers/ltr_trainer.py`.
+   Fixed on this branch.
+2. `lib/train/admin/local.py` (git-ignored, per-environment) pointed
+   `lasher_dir` / `depthtrack_dir` / `visevent_dir` one directory level too
+   shallow relative to how these datasets are actually laid out on this
+   machine. Corrected locally to:
+   - `lasher_dir` → `.../lasher/lasher/trainingset`
+   - `visevent_dir` → `.../visevent/visevent/train`
+   - `depthtrack_dir` → `.../depthtrack/depthtrack/train/DepthTrackTraining`
+
+---
+
+## ⚠️ SUPERSEDED: the "Analysis" section above uses a homegrown box-overlap metric
+
+The results table and analysis immediately above this line were computed with a
+**homegrown micro-averaged box-overlap AUC/PR** (`compute_ablation_metrics.py`),
+which is **not** the metric the paper/README actually reports. It produced a
+near-uniform ~7-8pt VisEvent gap vs the README across *every* config, including
+config-only ablations that barely touch the model — that was the tell that it was
+a metric-definition mismatch, not a real effect. Keep reading below for the
+corrected numbers. The relative ordering *within* that old table is still fine to
+reference, but do not quote its absolute numbers.
+
+## ✅ FINAL, OFFICIAL RESULTS (2026-07-07)
+
+Every number below is computed with the same tools the paper/README uses:
+- **RGBT234** → MPR/MSR via the official `rgbt` Python toolkit (`RGBT_toolkit_python/`, max-over-visible+infrared GT)
+- **LasHeR** → PR/SR via the same `rgbt` toolkit
+- **VisEvent** → OPE AUC/PR with absent-frame exclusion (`absent_label.txt`), 21 MATLAB-aligned thresholds, via `evaluate_lasher_visevent.py`
+- **DepthTrack** → official VOT toolkit (Pr/Re/F) — this one was *already* correct throughout
+
+Scoring script: `/mnt/task_runtime/rescore_official.py <config1> [<config2> ...]`, output written to
+`ablation_results_official/<config>/metrics.json`. Raw per-frame predictions live on each
+config's B200 node under `workspace/results/<Dataset[_miss]>/<config>/`.
+
+### Sanity check: does a fresh baseline reproduce the README?
+
+| | RGBT234 MSR | LasHeR SR | VisEvent AUC |
+|---|---:|---:|---:|
+| **README (published, tuned thresholds)** | 70.02 | 61.96 | 72.00 |
+| **rung-0, official metric, untuned thresholds** | 69.13–69.59 | 61.29–61.57 | 72.28–72.42 |
+| *(superseded)* rung-0, homegrown metric | ~66.8 | ~56.2 | ~63.7 |
+
+Yes — within ~1pt, fully explained by the README using per-dataset grid-searched
+test-time thresholds (UPT/UPH/INTER/MB) while every config below shares one
+untuned threshold set for a fair cross-config comparison.
+
+### Full table — 16 configs × precision-style + success-style metric × full/miss
+
+`P` = precision-style (MPR / PR / PR / Pr); `S` = success-style (MSR / SR / AUC / F).
+
+| config | RGBT234 full (P/S) | RGBT234 miss (P/S) | LasHeR full (P/S) | LasHeR miss (P/S) | VisEvent full (P/S) | VisEvent miss (P/S) | DepthTrack full (P/S) | DepthTrack miss (P/S) |
+|---|---|---|---|---|---|---|---|---|
+| rung0 (seed42) **[baseline]** | 91.94 / 69.13 | 84.55 / 62.66 | 76.33 / 61.29 | 67.63 / 54.41 | 88.18 / 72.33 | 78.06 / 62.20 | 65.55 / 67.01 | 56.61 / 57.88 |
+| rung0 (seed123) | 92.66 / 69.59 | 85.02 / 63.23 | 76.52 / 61.57 | 67.37 / 54.28 | 88.11 / 72.28 | 77.21 / 61.61 | 63.86 / 65.20 | 55.75 / 56.89 |
+| rung0 (seed456) | 91.52 / 69.29 | 84.69 / 62.89 | 76.67 / 61.37 | 66.42 / 53.42 | 88.08 / 72.42 | 77.04 / 61.61 | 61.66 / 62.95 | 56.58 / 57.83 |
+| **BMR mechanism family** | | | | | | | | |
+| no_hallucinate | 92.41 / 69.45 | 84.57 / 62.68 | 76.14 / 61.03 | 67.36 / 54.01 | 88.63 / 72.60 | 77.18 / 61.54 | 63.89 / 65.30 | 56.11 / 57.35 |
+| no_recon_loss | 91.57 / 68.94 | 85.20 / 62.98 | 76.28 / 61.16 | 66.13 / 53.35 | 87.48 / 72.06 | 77.20 / 61.52 | 63.93 / 65.30 | 54.81 / 56.01 |
+| uni_a2r | 92.27 / 69.63 | 84.34 / 62.68 | 75.98 / 60.94 | 66.81 / 53.84 | 87.23 / 71.66 | 76.80 / 61.35 | 65.13 / 66.53 | 58.32 / 59.53 |
+| uni_r2a | 92.23 / 69.15 | 83.93 / 62.31 | 76.03 / 60.89 | 67.27 / 54.15 | 88.20 / 72.48 | 76.78 / 61.35 | 63.27 / 64.62 | 57.53 / 58.74 |
+| no_ortho | 92.27 / 69.41 | 86.27 / 63.83 | 76.90 / 61.55 | 65.94 / 53.09 | 88.42 / 72.36 | 76.74 / 61.15 | 64.24 / 65.62 | 58.79 / 60.05 |
+| **MoE capacity / CMA / distill family** | | | | | | | | |
+| moe_small | 91.90 / 69.42 | 85.23 / 63.61 | 75.63 / 60.52 | 67.64 / 54.28 | 87.93 / 72.15 | 76.70 / 61.24 | 65.79 / 67.23 | 54.38 / 55.55 |
+| moe_middle | 91.90 / 69.14 | 85.81 / 63.44 | 77.02 / 61.74 | 67.70 / 54.46 | 87.58 / 72.03 | 76.59 / 61.19 | 60.95 / 62.30 | 54.73 / 55.95 |
+| moe_hybrid | 92.12 / 69.75 | 84.52 / 63.07 | 76.20 / 61.07 | 66.66 / 53.66 | 88.60 / 72.76 | 76.87 / 61.30 | 64.71 / 66.11 | 56.24 / 57.50 |
+| cma_fixed020 | 91.70 / 69.18 | 84.16 / 62.53 | 76.27 / 61.30 | 66.54 / 53.68 | 88.13 / 72.41 | 77.02 / 61.45 | 65.92 / 67.33 | 56.19 / 57.39 |
+| no_distill | 92.37 / 69.22 | 86.41 / 64.01 | 76.89 / 61.80 | 66.94 / 53.91 | 87.93 / 71.97 | 78.02 / 62.18 | 64.66 / 66.05 | 57.62 / 58.82 |
+| pmax_015 | 91.55 / 68.82 | 85.06 / 63.06 | 75.68 / 60.78 | 65.83 / 53.10 | 88.29 / 72.59 | 76.55 / 61.14 | 62.33 / 63.72 | 55.75 / 56.92 |
+| pmax_025 | 92.23 / 69.14 | 85.13 / 63.10 | 76.35 / 61.20 | 66.73 / 53.68 | 86.81 / 71.26 | 78.51 / 62.41 | 64.18 / 65.57 | 53.71 / 54.91 |
+| pmax_050 | 91.22 / 68.89 | 83.44 / 62.24 | 76.02 / 61.06 | 66.49 / 53.51 | 88.05 / 72.33 | 76.14 / 60.96 | 64.91 / 66.34 | 57.57 / 58.85 |
+
+Also published as an interactive web report:
+https://claude.ai/code/artifact/953bedd8-c304-43ee-8650-2ebde1646a27
+
+### What each ablation tests
+
+**BMR mechanism family** (isolates pieces of the bilateral hallucination/reconstruction design; config switches live under `cfg.MODEL.MOE` — `SUBSTITUTE_MODE`, `USE_RECON_LOSS`, `HALLUCINATE_DIRECTION`, `USE_ORTHO` in `lib/config/mcitrack/config.py`):
+- `no_hallucinate` (`SUBSTITUTE_MODE=zero`): missing modality is zero-filled instead of network-reconstructed. Does the hallucination network matter at all vs. naive zero-fill?
+- `no_recon_loss` (`USE_RECON_LOSS=false`): keeps the reconstruction network but drops the self-supervised reconstruction loss at train time. Is that auxiliary loss necessary, or does the architecture learn reconstruction anyway?
+- `uni_a2r` (`HALLUCINATE_DIRECTION=aux2rgb`): only the aux→rgb reconstruction path is active (rgb missing → reconstructed; aux missing → zero-filled). Tests whether one direction alone captures most of the benefit.
+- `uni_r2a` (`HALLUCINATE_DIRECTION=rgb2aux`): mirror of the above — only rgb→aux active.
+- `no_ortho` (`USE_ORTHO=false`): drops the MoE gate-orthogonality regularizer. Does forcing expert specialization actually help?
+
+**MoE capacity / CMA / distillation family** (the original Tier-1 batch, re-evaluated with the fixed pipeline):
+- `moe_small` / `moe_middle` / `moe_hybrid`: alternate expert-width configurations for the heterogeneous MoE (vs. the default 4→512 exponential spread). Tests sensitivity to expert capacity/heterogeneity.
+- `cma_fixed020`: cross-modal-attention modality-dropout probability held fixed at 0.20 during training instead of annealed 0.0→0.35. Tests whether the annealing curriculum matters vs. a fixed dropout rate.
+- `no_distill` (`DISTILL_WEIGHT=0`): removes the distillation loss entirely.
+- `pmax_015` / `pmax_025` / `pmax_050`: anneal ceiling (P_MAX) swept vs. the default 0.35. Tests sensitivity to how aggressively missing-modality is simulated during training.
+
+### Bugs found & fixed during this study (read before trusting any number above)
+
+1. **Missing-modality silent-fallback bug** — `test_rgbt_mgpus.py` / `test_depthtrack_mgpus.py` used to fall back to a dummy "always present" annotation whenever the real missing-modality JSON was absent on a node. This happened on **all 8 original B200 nodes** for the entire first Tier-1 batch (except the locally-run `moe_big`), making every `_miss` result in the *old, superseded* analysis above actually a full-modality run in disguise. Fixed: hard-fail (`raise FileNotFoundError`/`SystemExit`) instead of faking, JSON files shipped to every node.
+2. **VOT22RGBD wrong protocol assumed** — VOT-RGBD2022 uses the MultiStart/EAO live-trax protocol, not DepthTrack's precomputed-box Pr/Re/F protocol. Fixed the parser/harness (`run_vot22rgbd_local.sh`, `compute_ablation_metrics.py::eval_vot22rgbd`); **still unresolved**: live-trax evaluation hangs after ~2 sequences in this environment, root cause unknown. VOT22RGBD is excluded from all tables above.
+3. **Checkpoint scp truncation** — 2-minute foreground timeout truncated ~1.36GB checkpoint pulls. Fixed via background pulls with byte-count verification + retry.
+4. **Metric-definition mismatch (the big one)** — the homegrown box-overlap AUC/PR in `compute_ablation_metrics.py` is not what the paper's published table used. Fixed by re-scoring everything with `rescore_official.py` against the `rgbt` toolkit (RGBT234/LasHeR) and `evaluate_lasher_visevent.py` (VisEvent, absent-frame-aware OPE). This explains essentially all of the "-3 to -8pt gap vs README" seen in the old analysis.
+5. **Stale `metrics.json` carry-over** — when re-scoring the old 8 Tier-1 configs, DepthTrack numbers were carried over from a pre-fix `ablation_results/<config>/metrics.json` that hadn't been re-pulled from the node, producing a fake uniform ~66-67 DepthTrack_miss cluster. Caught by re-running `vot analysis --nocache` fresh on-node and comparing; fixed by re-pulling and patching all 8 configs' metrics files.
+
+### Known limitation flagged to the user (2026-07-07)
+
+Only rung-0 has 3 seeds. All 13 ablations above are **single-seed** — "outside the
+rung-0 noise band" is suggestive, not confirmed, until the ablations themselves are
+repeated across seeds. None of the 5 BMR-mechanism ablations currently clears a
+strong bar above noise on missing-modality metrics.
+
+### In progress: missing-ratio sweep (requested 2026-07-07)
+
+Single-point `_miss` eval (one fixed official missing pattern) doesn't show much
+separation between mechanism ablations. Building a synthetic missing-**ratio**
+sweep (0/25/50/75/100% of frames with the aux modality zeroed) to plot
+performance-vs-missing-rate curves instead of a single point — the expectation is
+that `no_hallucinate` and friends should diverge from rung-0 more visibly at high
+missing rates. Uses the same already-trained checkpoints (no retraining needed),
+inference-only. See below for the generator + sweep plan/status.

ablation_results/README.md

@@ -0,0 +1,37 @@
+# FlexTrack-V2 (V56) ablation study — results package
+
+Full ablation study for the BMR-HMoE mechanism and MoE/CMA/distillation
+hyperparameters, scored with the same official toolkits the paper uses
+(rgbt toolkit for RGBT234/LasHeR, absent-frame-aware OPE for VisEvent, VOT
+toolkit for DepthTrack). See `ABLATIONS.md` for the full narrative: what each
+ablation tests, every bug found and fixed along the way, and the final
+16-config results table. `official_ablation_report.html` is the same content
+as an interactive page (open directly in a browser).
+
+## Contents
+
+- `ABLATIONS.md` — full writeup: ablation design, bugs found/fixed, final table
+- `official_ablation_report.html` — interactive version of the results table
+- `official_metrics/<config>/metrics.json` — final official-toolkit scores
+  for all 16 ablation configs (RGBT234/LasHeR/VisEvent full+miss via
+  rgbt-toolkit/OPE, DepthTrack full+miss via VOT)
+- `raw_metrics/<config>/metrics.json` — earlier-stage per-config metrics
+  (includes `mcitrack_b224_56_abl_moe_big` and the large-backbone
+  `mcitrack_l224_56`, both evaluated separately from the main 16-config set)
+- `grid_search_final_params/` — the coordinate-descent grid search logs'
+  final params for the base model's test-time thresholds (UPT/UPH/INTER/MB).
+  RGBT234 (full+miss), LasHeR (full), and DepthTrack (full+miss) converged
+  properly (explicit "steps at floor, stopping search" termination). VisEvent
+  full's search log ends mid-round with no convergence marker — its params
+  are the best found, not a certified optimum. LasHeR_miss/VisEvent_miss have
+  no grid search on record.
+- `vot22rgbd_status/` — VOT22RGBD produced no usable data (live-trax hang);
+  documented here rather than silently omitted.
+
+## Caveat shared by every ablation config
+
+All 16 ablation configs (including rung-0) share one common, untuned
+test-time threshold set for a fair cross-config comparison — none of them
+use the grid-search-tuned thresholds above. That means absolute numbers here
+run ~0.1-1pt under the tuned headline numbers; the *relative* comparison
+across configs is what's meaningful.

ablation_results/grid_search_final_params/gs_depthtrack_full_v2_FINAL.json

@@ -0,0 +1,23 @@
+{
+  "round": 4,
+  "params": {
+    "upt": 0.77,
+    "uph": 0.85,
+    "inter": 55,
+    "mb": 862
+  },
+  "metrics": {
+    "Pr": 66.42876272739738,
+    "Re": 69.28717754185934,
+    "F": 67.82786852963147
+  },
+  "avg": 67.84793626629606,
+  "target": {
+    "Pr": 67.1,
+    "Re": 66.9,
+    "F": 67.0
+  },
+  "target_avg": 67.0,
+  "beats_flextrack_on_every_metric": false,
+  "beats_flextrack_on_average": true
+}

ablation_results/grid_search_final_params/gs_depthtrack_miss_v2_FINAL.json

@@ -0,0 +1,23 @@
+{
+  "round": 4,
+  "params": {
+    "upt": 0.77,
+    "uph": 0.81,
+    "inter": 55,
+    "mb": 517
+  },
+  "metrics": {
+    "Pr": 57.53250451599064,
+    "Re": 60.098073643583035,
+    "F": 58.78731104443481
+  },
+  "avg": 58.805963068002825,
+  "target": {
+    "Pr": 59.6,
+    "Re": 56.1,
+    "F": 57.8
+  },
+  "target_avg": 57.833333333333336,
+  "beats_flextrack_on_every_metric": false,
+  "beats_flextrack_on_average": true
+}

ablation_results/grid_search_final_params/gs_lasher_full_FINAL.json

@@ -0,0 +1,17 @@
+{
+  "params": {
+    "UPT": 0.6,
+    "UPH": 0.98,
+    "INTER": 10,
+    "MB": 100
+  },
+  "metrics": {
+    "pr": 77.32249784086656,
+    "auc": 61.95738421617586
+  },
+  "targets": {
+    "pr": 77.28,
+    "auc": 61.94
+  },
+  "met": true
+}

ablation_results/grid_search_final_params/gs_rgbt234_full_FINAL.json

@@ -0,0 +1,16 @@
+{
+  "params": {
+    "UPT": 0.5,
+    "UPH": 0.93,
+    "INTER": 25,
+    "MB": 312
+  },
+  "metrics": {
+    "mpr": 93.37423905144344,
+    "msr": 70.01662575965612
+  },
+  "targets": {
+    "mpr": 92.72
+  },
+  "met": true
+}

ablation_results/grid_search_final_params/gs_rgbt234_miss_FINAL.json

@@ -0,0 +1,17 @@
+{
+  "params": {
+    "UPT": 0.35,
+    "UPH": 0.88,
+    "INTER": 18,
+    "MB": 250
+  },
+  "metrics": {
+    "mpr": 85.70714894107877,
+    "msr": 62.804693653433134
+  },
+  "targets": {
+    "mpr": 84.07,
+    "msr": 62.73
+  },
+  "met": true
+}

ablation_results/grid_search_final_params/gs_visevent_full_FINAL.json

@@ -0,0 +1,17 @@
+{
+  "params": {
+    "UPT": 0.8,
+    "UPH": 0.98,
+    "INTER": 70,
+    "MB": 300
+  },
+  "metrics": {
+    "auc": 71.65672285984303,
+    "pr": 87.46859988541713
+  },
+  "targets": {
+    "auc": 72.0,
+    "pr": 88.0
+  },
+  "met": false
+}

ablation_results/grid_search_final_params/gs_visevent_full_bonus_FINAL.json

@@ -0,0 +1,18 @@
+{
+  "round": 2,
+  "params": {
+    "UPT": 0.7,
+    "UPH": 0.98,
+    "INTER": 70,
+    "MB": 100
+  },
+  "metrics": {
+    "auc": 71.99648416650753,
+    "pr": 87.92252434886079
+  },
+  "targets": {
+    "auc": 72.16,
+    "pr": 87.97
+  },
+  "met": false
+}

ablation_results/official_ablation_report.html

@@ -0,0 +1,329 @@
+<title>FlexTrack-V2 · 16 Ablations, Official Metrics</title>
+<style>
+  :root{
+    --paper:#eef2f5; --raised:#ffffff; --ink:#14202b; --slate:#576775; --slate-2:#8494a1;
+    --line:#d6dee5; --line-2:#c0cbd4; --accent:#0e7c86; --accent-ink:#0a545b; --accent-soft:#dceff0;
+    --band:#c9d6de; --band-soft:#e2e9ee; --warn:#b5532b; --warn-soft:#f6e7df; --good:#3c7d57; --good-soft:#e1efe6;
+    --serif:Charter,"Iowan Old Style","Palatino Linotype",Georgia,"Times New Roman",serif;
+    --sans:system-ui,-apple-system,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;
+    --mono:ui-monospace,"SF Mono","Cascadia Code","Roboto Mono",Menlo,Consolas,monospace;
+  }
+  @media (prefers-color-scheme:dark){:root{
+    --paper:#0d141a; --raised:#151f27; --ink:#e6edf2; --slate:#9fb0bc; --slate-2:#6d7e8b;
+    --line:#26333d; --line-2:#374855; --accent:#3fb6c2; --accent-ink:#a8e4ea; --accent-soft:#122b2f;
+    --band:#33434f; --band-soft:#1b2932; --warn:#e07a52; --warn-soft:#2c1c14; --good:#5cb381; --good-soft:#132821;
+  }}
+  :root[data-theme="light"]{
+    --paper:#eef2f5; --raised:#ffffff; --ink:#14202b; --slate:#576775; --slate-2:#8494a1;
+    --line:#d6dee5; --line-2:#c0cbd4; --accent:#0e7c86; --accent-ink:#0a545b; --accent-soft:#dceff0;
+    --band:#c9d6de; --band-soft:#e2e9ee; --warn:#b5532b; --warn-soft:#f6e7df; --good:#3c7d57; --good-soft:#e1efe6;
+  }
+  :root[data-theme="dark"]{
+    --paper:#0d141a; --raised:#151f27; --ink:#e6edf2; --slate:#9fb0bc; --slate-2:#6d7e8b;
+    --line:#26333d; --line-2:#374855; --accent:#3fb6c2; --accent-ink:#a8e4ea; --accent-soft:#122b2f;
+    --band:#33434f; --band-soft:#1b2932; --warn:#e07a52; --warn-soft:#2c1c14; --good:#5cb381; --good-soft:#132821;
+  }
+  *{box-sizing:border-box}
+  body{margin:0;background:var(--paper);color:var(--ink);font-family:var(--sans);line-height:1.55;
+    padding:0 0 5rem;-webkit-font-smoothing:antialiased}
+  ::selection{background:var(--accent-soft);color:var(--accent-ink)}
+  .wrap{max-width:1180px;margin:0 auto;padding:3.2rem 1.6rem 0}
+  .eyebrow{font-family:var(--mono);font-size:.74rem;letter-spacing:.13em;text-transform:uppercase;color:var(--accent)}
+  h1{font-family:var(--serif);font-weight:600;font-size:clamp(1.85rem,3.4vw,2.6rem);line-height:1.12;
+    margin:.7rem 0 .5rem;letter-spacing:-.015em;text-wrap:balance}
+  .sub{font-size:1.05rem;color:var(--slate);max-width:68ch;text-wrap:pretty}
+  header{border-bottom:1px solid var(--line);padding-bottom:2rem}
+  .meta{display:flex;flex-wrap:wrap;gap:.5rem;margin-top:1.2rem}
+  .chip{font-family:var(--mono);font-size:.72rem;letter-spacing:.02em;color:var(--slate);
+    background:var(--raised);border:1px solid var(--line);border-radius:999px;padding:.28rem .7rem}
+  .chip b{color:var(--ink);font-weight:600}
+  section{margin-top:2.6rem}
+  h2{font-family:var(--serif);font-size:1.45rem;font-weight:600;margin:0 0 .3rem;letter-spacing:-.01em;text-wrap:balance}
+  .lead{color:var(--slate);max-width:76ch;margin:.4rem 0 1rem}
+  h3{font-size:.98rem;font-weight:700;margin:1.2rem 0 .4rem}
+
+  .fixbox{background:var(--warn-soft);border:1px solid color-mix(in srgb,var(--warn) 30%,var(--line));
+    border-radius:12px;padding:1.1rem 1.3rem;margin-top:1rem}
+  .fixbox .k{font-family:var(--mono);font-size:.72rem;letter-spacing:.05em;text-transform:uppercase;color:var(--warn);margin-bottom:.5rem}
+  .fixbox p{margin:.35rem 0;color:var(--ink)}
+
+  .compare3{display:grid;grid-template-columns:repeat(3,1fr);gap:1px;background:var(--line);
+    border:1px solid var(--line);border-radius:9px;overflow:hidden;margin-top:1rem}
+  .compare3>div{background:var(--raised);padding:.85rem 1rem}
+  .compare3 .hd{font-size:.7rem;text-transform:uppercase;letter-spacing:.05em;color:var(--slate);margin-bottom:.5rem}
+  .compare3 .row{display:flex;justify-content:space-between;font-family:var(--mono);font-size:.85rem;padding:.15rem 0}
+  .compare3 .row .k{color:var(--slate);font-family:var(--sans)}
+  .compare3 .match{color:var(--good)}
+
+  .scroll{overflow-x:auto;border:1px solid var(--line);border-radius:12px;margin-top:.8rem;background:var(--raised)}
+  table{border-collapse:collapse;width:100%;min-width:1560px;font-size:.8rem}
+  thead th{position:sticky;top:0;background:var(--raised);text-align:right;padding:.55rem .65rem;border-bottom:2px solid var(--line-2);
+    font-weight:700;white-space:nowrap;font-size:.75rem}
+  thead th:first-child{text-align:left}
+  thead .grp{font-family:var(--mono);font-size:.66rem;color:var(--slate);font-weight:600;letter-spacing:.03em}
+  tbody td{padding:.42rem .65rem;text-align:right;font-family:var(--mono);font-variant-numeric:tabular-nums;
+    border-bottom:1px solid var(--line);white-space:nowrap}
+  tbody td:first-child{text-align:left;font-family:var(--sans);font-weight:600}
+  tbody tr.base td{background:var(--band-soft)}
+  tbody tr.tunedrow td{background:var(--accent-soft);font-weight:600}
+  tbody tr.tunedrow td .fam{font-weight:400}
+  tbody tr.baseend td{border-bottom:2px solid var(--line-2)}
+  tbody tr.grpstart td{border-top:2px solid var(--line-2)}
+  tbody tr:last-child td{border-bottom:none}
+  .miss{color:var(--slate)}
+  .fam{font-family:var(--mono);font-size:.64rem;color:var(--slate-2);text-transform:uppercase;letter-spacing:.03em}
+
+  .bands{display:flex;flex-direction:column;gap:.1rem;margin-top:.8rem;border:1px solid var(--line);border-radius:12px;
+    background:var(--raised);padding:1rem 1.1rem}
+  .brow{display:grid;grid-template-columns:150px 1fr 130px;align-items:center;gap:.8rem;padding:.45rem 0}
+  .brow+.brow{border-top:1px solid var(--line)}
+  .brow .lbl{font-family:var(--mono);font-size:.76rem;color:var(--ink)}
+  .track{position:relative;height:26px}
+  .axis{position:absolute;left:0;right:0;top:50%;height:1px;background:var(--line-2)}
+  .bandbox{position:absolute;top:5px;bottom:5px;background:var(--band);opacity:.55;border-radius:3px}
+  .dot{position:absolute;top:50%;width:9px;height:9px;border-radius:50%;transform:translate(-50%,-50%);border:1.5px solid var(--raised)}
+  .dot.in{background:var(--accent)} .dot.out{background:var(--warn)}
+  .brow .verdict{font-family:var(--mono);font-size:.7rem;text-align:right;color:var(--slate)}
+  .legend{display:flex;flex-wrap:wrap;gap:1.1rem;margin-top:.7rem;font-size:.75rem;color:var(--slate);align-items:center}
+  .legend span{display:inline-flex;align-items:center;gap:.4rem}
+  .sw{width:11px;height:11px;border-radius:3px;display:inline-block}
+  .sw.band{background:var(--band)} .sw.din{background:var(--accent);border-radius:50%} .sw.dout{background:var(--warn);border-radius:50%}
+
+  .findings{display:flex;flex-direction:column;gap:.85rem;margin-top:1rem}
+  .f{display:grid;grid-template-columns:2rem 1fr;gap:.85rem;align-items:start}
+  .f .n{width:2rem;height:2rem;border-radius:7px;display:flex;align-items:center;justify-content:center;
+    font-family:var(--mono);font-weight:700;font-size:.88rem;background:var(--accent-soft);color:var(--accent-ink)}
+  .f.warn .n{background:var(--warn-soft);color:var(--warn)}
+  .f.good .n{background:var(--good-soft);color:var(--good)}
+  .f h4{margin:.1rem 0 .2rem;font-size:.98rem}
+  .f p{margin:0;color:var(--slate);font-size:.92rem}
+  .caveat{border-left:3px solid var(--line-2);padding:.15rem 0 .15rem 1rem;color:var(--slate);font-size:.88rem;margin-top:.9rem}
+  code{font-family:var(--mono);font-size:.85em;background:var(--band-soft);padding:.05rem .35rem;border-radius:4px}
+  footer{margin-top:3rem;padding-top:1.2rem;border-top:1px solid var(--line);font-family:var(--mono);font-size:.74rem;color:var(--slate-2)}
+  @media(max-width:720px){.compare3,.brow{grid-template-columns:1fr}.wrap{padding:2.2rem 1.05rem 0}}
+</style>
+
+<div class="wrap">
+  <header>
+    <div class="eyebrow">FlexTrack-V2 (V56) · full ablation study, official protocols</div>
+    <h1>16 configs, scored with your own paper's evaluators</h1>
+    <p class="sub">Every prediction re-scored with the <em>official</em> toolkit per dataset — <code>rgbt</code> package for RGBT234 (MPR/MSR) and LasHeR (PR/SR), OPE with absent-frame handling for VisEvent, and the VOT toolkit for DepthTrack — the same code that produced the paper's headline numbers. This replaces an earlier pass that used a homegrown box-overlap metric.</p>
+    <div class="meta">
+      <span class="chip"><b>16</b> configs · <b>8</b> official metric columns</span>
+      <span class="chip">rung-0 × <b>3 seeds</b> — the noise floor</span>
+      <span class="chip">2 ablation families: <b>BMR mechanism</b> · <b>MoE / CMA / distill</b></span>
+    </div>
+  </header>
+
+  <section>
+    <h2>Reproducing the paper number, first</h2>
+    <p class="lead">Before trusting any ablation, the official evaluator has to reproduce your published baseline on a freshly-trained "rung-0" (identical recipe, seed 42). It does — the earlier ~8 point VisEvent gap was <em>entirely</em> a metric artifact, not a real ablation effect or a training problem.</p>
+    <div class="compare3">
+      <div><div class="hd">Published (README)</div>
+        <div class="row"><span class="k">VisEvent AUC</span><span>72.00</span></div>
+        <div class="row"><span class="k">RGBT234 MSR</span><span>70.02</span></div>
+        <div class="row"><span class="k">LasHeR SR</span><span>61.96</span></div>
+      </div>
+      <div><div class="hd">rung-0, official tools</div>
+        <div class="row"><span class="k">VisEvent AUC</span><span class="match">72.33</span></div>
+        <div class="row"><span class="k">RGBT234 MSR</span><span class="match">69.13–69.59</span></div>
+        <div class="row"><span class="k">LasHeR SR</span><span class="match">61.29–61.57</span></div>
+      </div>
+      <div><div class="hd">rung-0, homegrown metric (superseded)</div>
+        <div class="row"><span class="k">VisEvent AUC</span><span>63.7</span></div>
+        <div class="row"><span class="k">RGBT234 MSR</span><span>66.8</span></div>
+        <div class="row"><span class="k">LasHeR SR</span><span>56.2</span></div>
+      </div>
+    </div>
+    <p class="caveat">Small residuals (RGBT234/LasHeR ≈1 pt) remain because the published table also used per-dataset grid-searched test-time thresholds; every config below shares one untuned threshold set for a fair comparison, so absolute values here run slightly under the tuned headline numbers.</p>
+  </section>
+
+  <section>
+    <h2>Base model, fully tuned vs. ablation-study default thresholds</h2>
+    <p class="lead">Every one of the 16 ablation configs below shares one untuned threshold set (UPT/UPH/INTER/MB), by design, for a fair apples-to-apples comparison across configs. That means every number in the big table understates what the base model can actually do once its test-time thresholds are grid-searched per dataset. This section puts both side by side, dataset by dataset, so the gap is explicit rather than a footnote.</p>
+    <div class="scroll"><table id="tunedtbl"></table></div>
+    <p class="caveat"><b>VisEvent full is flagged, not a clean number.</b> Its grid search log ends mid-round-3 with no convergence marker (no "steps at floor" line like every other dataset has) — 71.996 AUC is the best value the search happened to find before it was cut off, not a certified optimum. Notably it's <em>lower</em> than rung-0's own default-threshold 72.33, which is itself a sign the search never properly converged for this dataset.</p>
+    <p class="caveat"><b>LasHeR_miss and VisEvent_miss have no tuned column</b> — no grid search was ever run for these two; the "tuned" cell repeats the shared default since there's nothing else to show.</p>
+  </section>
+
+  <section>
+    <h2>A bug caught mid-analysis, for transparency</h2>
+    <div class="fixbox">
+      <div class="k">Correction applied before this table</div>
+      <p>The first pass of this table showed all 8 "MoE/CMA/distill" configs clustered at DepthTrack_miss F≈62–67, far outside the rung-0 band (57) — and identical to numbers from an earlier, since-corrected run. Re-running <code>vot analysis</code> fresh confirmed the true value (e.g. <code>cma_fixed020</code>: 57.39, not 66.11): a stale <code>metrics.json</code> had been carried over instead of re-pulled after re-evaluation. Fixed by re-fetching all 8 post-eval result files before scoring. Flagging this precisely because the same silent-staleness failure mode is exactly what caused the original missing-modality bug.</p>
+    </div>
+  </section>
+
+  <section>
+    <h2>Full results — 16 configs × precision + success metrics, full &amp; missing</h2>
+    <p class="lead">Every dataset shows both its precision-style metric (MPR / PR / PR / Pr) and its success/overlap-style metric (MSR / SR / AUC / F), for both full-modality and missing-modality test settings — 16 numeric columns total, all from the official evaluators. Shaded rows = rung-0 baseline (3 seeds); rows below are the BMR-mechanism family (5) then the MoE/CMA/distillation family (8).</p>
+    <div class="scroll"><table id="tbl"></table></div>
+  </section>
+
+  <section>
+    <h2>Is any ablation outside the noise band?</h2>
+    <p class="lead">Grey band = rung-0 3-seed min–max on each missing-modality metric. Each dot is one ablation (13 total, both families combined). A dot inside the band means, on this checkpoint, that component/change made no difference distinguishable from re-training with a different seed.</p>
+    <div class="bands" id="bands"></div>
+    <div class="legend">
+      <span><span class="sw band"></span> rung-0 3-seed band</span>
+      <span><span class="sw din"></span> inside (no effect)</span>
+      <span><span class="sw dout"></span> outside band</span>
+    </div>
+  </section>
+
+  <section>
+    <h2>What it means</h2>
+    <div class="findings">
+      <div class="f good"><div class="n">✓</div><div>
+        <h4>The evaluation pipeline is now trustworthy</h4>
+        <p>rung-0 reproduces the paper's numbers under official scoring; the missing-modality dummy-fallback bug is fixed and verified (RGBT234_miss dropped from a fake ≈full to a real ~9pt gap); the stale-metrics bug caught above is fixed too. Every number in the table above was scored by the paper's own evaluators.</p>
+      </div></div>
+      <div class="f"><div class="n">1</div><div>
+        <h4>Most single-run deltas are still inside or barely outside a narrow noise band</h4>
+        <p>Some columns have a genuinely tight band (VisEvent AUC ±0.07, RGBT234 MSR ±0.23) — where almost any single independent run will land "outside" just from sampling one point instead of three. Others (DepthTrack full ±2, LasHeR_miss ±0.5) give more room. Treat "outside band" here as <em>worth a second seed</em>, not as a confirmed effect — none of the 13 ablations show a deviation large relative to DepthTrack's own full-modality seed spread (62.95→67.01).</p>
+      </div></div>
+      <div class="f warn"><div class="n">!</div><div>
+        <h4>No mechanism ablation clears a high bar for "the component matters"</h4>
+        <p><code>no_hallucinate</code> (zero-fill vs. reconstruct — the core BMR claim), <code>no_recon_loss</code>, <code>uni_a2r</code>/<code>uni_r2a</code>, and <code>no_ortho</code> all land within about a point of rung-0 on every missing-modality metric. If the paper's narrative rests on these specific components driving the missing-modality gains, this single-seed pass doesn't yet demonstrate it — multi-seed repeats of the ablations (not just the baseline) are the natural next step before drawing conclusions either way.</p>
+      </div></div>
+    </div>
+  </section>
+
+  <section>
+    <h2>Status &amp; caveats</h2>
+    <p class="caveat"><b>Mechanism and MoE/CMA ablations are single-seed;</b> only rung-0 has 3. A config sitting inside the band is evidence of "no effect beyond noise" for that one run, not a confirmed null across seeds.</p>
+    <p class="caveat"><b>VOT22RGBD (a 9th, separate setting) isn't in this table.</b> Its VOT MultiStart/EAO live-trax evaluation hangs after ~2 sequences in this environment; unresolved, doesn't affect the 8 metrics above.</p>
+    <p class="caveat"><b>Absolute values sit ~1 pt under the tuned headline numbers</b> by design — every config here shares one untuned threshold set, so the comparison across configs is fair even though no row is individually optimized.</p>
+  </section>
+
+  <footer>ablation_results_official/*/metrics.json · rgbt toolkit (MPR/MSR, PR/SR) + evaluate_lasher_visevent.py (OPE) + VOT · generated 2026-07-07</footer>
+</div>
+
+<script>
+const DATA = {"base_tuned": {"config": "mcitrack_b224_56 (grid-search tuned, per-dataset thresholds)", "RGBT234": {"mpr": 93.37, "msr": 70.02}, "RGBT234_miss": {"mpr": 85.71, "msr": 62.80}, "LasHeR": {"pr": 77.32, "sr": 61.96}, "LasHeR_miss": null, "VisEvent": {"auc": 71.996, "pr": 87.92}, "VisEvent_miss": null, "DepthTrack": {"precision": 66.43, "recall": 69.29, "fscore": 67.83}, "DepthTrack_miss": {"precision": 57.53, "recall": 60.10, "fscore": 58.79}}, "rung0": {"config": "mcitrack_b224_56_abl_rung0", "RGBT234": {"mpr": 91.94, "msr": 69.13}, "RGBT234_miss": {"mpr": 84.55, "msr": 62.66}, "LasHeR": {"pr": 76.33, "sr": 61.29}, "LasHeR_miss": {"pr": 67.63, "sr": 54.41}, "VisEvent": {"auc": 72.33, "pr": 88.18}, "VisEvent_miss": {"auc": 62.2, "pr": 78.06}, "DepthTrack": {"precision": 65.55, "recall": 68.52, "fscore": 67.01}, "DepthTrack_miss": {"precision": 56.61, "recall": 59.21, "fscore": 57.88}}, "rung0_seedB": {"config": "mcitrack_b224_56_abl_rung0_seedB", "RGBT234": {"mpr": 92.66, "msr": 69.59}, "RGBT234_miss": {"mpr": 85.02, "msr": 63.23}, "LasHeR": {"pr": 76.52, "sr": 61.57}, "LasHeR_miss": {"pr": 67.37, "sr": 54.28}, "VisEvent": {"auc": 72.28, "pr": 88.11}, "VisEvent_miss": {"auc": 61.61, "pr": 77.21}, "DepthTrack": {"precision": 63.86, "recall": 66.61, "fscore": 65.2}, "DepthTrack_miss": {"precision": 55.75, "recall": 58.07, "fscore": 56.89}}, "rung0_seedC": {"config": "mcitrack_b224_56_abl_rung0_seedC", "RGBT234": {"mpr": 91.52, "msr": 69.29}, "RGBT234_miss": {"mpr": 84.69, "msr": 62.89}, "LasHeR": {"pr": 76.67, "sr": 61.37}, "LasHeR_miss": {"pr": 66.42, "sr": 53.42}, "VisEvent": {"auc": 72.42, "pr": 88.08}, "VisEvent_miss": {"auc": 61.61, "pr": 77.04}, "DepthTrack": {"precision": 61.66, "recall": 64.29, "fscore": 62.95}, "DepthTrack_miss": {"precision": 56.58, "recall": 59.14, "fscore": 57.83}}, "no_hallucinate": {"config": "mcitrack_b224_56_abl_no_hallucinate", "RGBT234": {"mpr": 92.41, "msr": 69.45}, "RGBT234_miss": {"mpr": 84.57, "msr": 62.68}, "LasHeR": {"pr": 76.14, "sr": 61.03}, "LasHeR_miss": {"pr": 67.36, "sr": 54.01}, "VisEvent": {"auc": 72.6, "pr": 88.63}, "VisEvent_miss": {"auc": 61.54, "pr": 77.18}, "DepthTrack": {"precision": 63.89, "recall": 66.77, "fscore": 65.3}, "DepthTrack_miss": {"precision": 56.11, "recall": 58.64, "fscore": 57.35}}, "no_recon_loss": {"config": "mcitrack_b224_56_abl_no_recon_loss", "RGBT234": {"mpr": 91.57, "msr": 68.94}, "RGBT234_miss": {"mpr": 85.2, "msr": 62.98}, "LasHeR": {"pr": 76.28, "sr": 61.16}, "LasHeR_miss": {"pr": 66.13, "sr": 53.35}, "VisEvent": {"auc": 72.06, "pr": 87.48}, "VisEvent_miss": {"auc": 61.52, "pr": 77.2}, "DepthTrack": {"precision": 63.93, "recall": 66.74, "fscore": 65.3}, "DepthTrack_miss": {"precision": 54.81, "recall": 57.27, "fscore": 56.01}}, "uni_a2r": {"config": "mcitrack_b224_56_abl_uni_a2r", "RGBT234": {"mpr": 92.27, "msr": 69.63}, "RGBT234_miss": {"mpr": 84.34, "msr": 62.68}, "LasHeR": {"pr": 75.98, "sr": 60.94}, "LasHeR_miss": {"pr": 66.81, "sr": 53.84}, "VisEvent": {"auc": 71.66, "pr": 87.23}, "VisEvent_miss": {"auc": 61.35, "pr": 76.8}, "DepthTrack": {"precision": 65.13, "recall": 67.99, "fscore": 66.53}, "DepthTrack_miss": {"precision": 58.32, "recall": 60.78, "fscore": 59.53}}, "uni_r2a": {"config": "mcitrack_b224_56_abl_uni_r2a", "RGBT234": {"mpr": 92.23, "msr": 69.15}, "RGBT234_miss": {"mpr": 83.93, "msr": 62.31}, "LasHeR": {"pr": 76.03, "sr": 60.89}, "LasHeR_miss": {"pr": 67.27, "sr": 54.15}, "VisEvent": {"auc": 72.48, "pr": 88.2}, "VisEvent_miss": {"auc": 61.35, "pr": 76.78}, "DepthTrack": {"precision": 63.27, "recall": 66.04, "fscore": 64.62}, "DepthTrack_miss": {"precision": 57.53, "recall": 60.0, "fscore": 58.74}}, "no_ortho": {"config": "mcitrack_b224_56_abl_no_ortho", "RGBT234": {"mpr": 92.27, "msr": 69.41}, "RGBT234_miss": {"mpr": 86.27, "msr": 63.83}, "LasHeR": {"pr": 76.9, "sr": 61.55}, "LasHeR_miss": {"pr": 65.94, "sr": 53.09}, "VisEvent": {"auc": 72.36, "pr": 88.42}, "VisEvent_miss": {"auc": 61.15, "pr": 76.74}, "DepthTrack": {"precision": 64.24, "recall": 67.06, "fscore": 65.62}, "DepthTrack_miss": {"precision": 58.79, "recall": 61.36, "fscore": 60.05}}, "moe_small": {"config": "mcitrack_b224_56_abl_moe_small", "RGBT234": {"mpr": 91.9, "msr": 69.42}, "RGBT234_miss": {"mpr": 85.23, "msr": 63.61}, "LasHeR": {"pr": 75.63, "sr": 60.52}, "LasHeR_miss": {"pr": 67.64, "sr": 54.28}, "VisEvent": {"auc": 72.15, "pr": 87.93}, "VisEvent_miss": {"auc": 61.24, "pr": 76.7}, "DepthTrack": {"precision": 65.79, "recall": 68.74, "fscore": 67.23}, "DepthTrack_miss": {"precision": 54.38, "recall": 56.77, "fscore": 55.55}}, "moe_middle": {"config": "mcitrack_b224_56_abl_moe_middle", "RGBT234": {"mpr": 91.9, "msr": 69.14}, "RGBT234_miss": {"mpr": 85.81, "msr": 63.44}, "LasHeR": {"pr": 77.02, "sr": 61.74}, "LasHeR_miss": {"pr": 67.7, "sr": 54.46}, "VisEvent": {"auc": 72.03, "pr": 87.58}, "VisEvent_miss": {"auc": 61.19, "pr": 76.59}, "DepthTrack": {"precision": 60.95, "recall": 63.7, "fscore": 62.3}, "DepthTrack_miss": {"precision": 54.73, "recall": 57.22, "fscore": 55.95}}, "moe_hybrid": {"config": "mcitrack_b224_56_abl_moe_hybrid", "RGBT234": {"mpr": 92.12, "msr": 69.75}, "RGBT234_miss": {"mpr": 84.52, "msr": 63.07}, "LasHeR": {"pr": 76.2, "sr": 61.07}, "LasHeR_miss": {"pr": 66.66, "sr": 53.66}, "VisEvent": {"auc": 72.76, "pr": 88.6}, "VisEvent_miss": {"auc": 61.3, "pr": 76.87}, "DepthTrack": {"precision": 64.71, "recall": 67.56, "fscore": 66.11}, "DepthTrack_miss": {"precision": 56.24, "recall": 58.81, "fscore": 57.5}}, "cma_fixed020": {"config": "mcitrack_b224_56_abl_cma_fixed020", "RGBT234": {"mpr": 91.7, "msr": 69.18}, "RGBT234_miss": {"mpr": 84.16, "msr": 62.53}, "LasHeR": {"pr": 76.27, "sr": 61.3}, "LasHeR_miss": {"pr": 66.54, "sr": 53.68}, "VisEvent": {"auc": 72.41, "pr": 88.13}, "VisEvent_miss": {"auc": 61.45, "pr": 77.02}, "DepthTrack": {"precision": 65.92, "recall": 68.8, "fscore": 67.33}, "DepthTrack_miss": {"precision": 56.19, "recall": 58.64, "fscore": 57.39}}, "no_distill": {"config": "mcitrack_b224_56_abl_no_distill", "RGBT234": {"mpr": 92.37, "msr": 69.22}, "RGBT234_miss": {"mpr": 86.41, "msr": 64.01}, "LasHeR": {"pr": 76.89, "sr": 61.8}, "LasHeR_miss": {"pr": 66.94, "sr": 53.91}, "VisEvent": {"auc": 71.97, "pr": 87.93}, "VisEvent_miss": {"auc": 62.18, "pr": 78.02}, "DepthTrack": {"precision": 64.66, "recall": 67.5, "fscore": 66.05}, "DepthTrack_miss": {"precision": 57.62, "recall": 60.07, "fscore": 58.82}}, "pmax_015": {"config": "mcitrack_b224_56_abl_pmax_015", "RGBT234": {"mpr": 91.55, "msr": 68.82}, "RGBT234_miss": {"mpr": 85.06, "msr": 63.06}, "LasHeR": {"pr": 75.68, "sr": 60.78}, "LasHeR_miss": {"pr": 65.83, "sr": 53.1}, "VisEvent": {"auc": 72.59, "pr": 88.29}, "VisEvent_miss": {"auc": 61.14, "pr": 76.55}, "DepthTrack": {"precision": 62.33, "recall": 65.18, "fscore": 63.72}, "DepthTrack_miss": {"precision": 55.75, "recall": 58.15, "fscore": 56.92}}, "pmax_025": {"config": "mcitrack_b224_56_abl_pmax_025", "RGBT234": {"mpr": 92.23, "msr": 69.14}, "RGBT234_miss": {"mpr": 85.13, "msr": 63.1}, "LasHeR": {"pr": 76.35, "sr": 61.2}, "LasHeR_miss": {"pr": 66.73, "sr": 53.68}, "VisEvent": {"auc": 71.26, "pr": 86.81}, "VisEvent_miss": {"auc": 62.41, "pr": 78.51}, "DepthTrack": {"precision": 64.18, "recall": 67.02, "fscore": 65.57}, "DepthTrack_miss": {"precision": 53.71, "recall": 56.16, "fscore": 54.91}}, "pmax_050": {"config": "mcitrack_b224_56_abl_pmax_050", "RGBT234": {"mpr": 91.22, "msr": 68.89}, "RGBT234_miss": {"mpr": 83.44, "msr": 62.24}, "LasHeR": {"pr": 76.02, "sr": 61.06}, "LasHeR_miss": {"pr": 66.49, "sr": 53.51}, "VisEvent": {"auc": 72.33, "pr": 88.05}, "VisEvent_miss": {"auc": 60.96, "pr": 76.14}, "DepthTrack": {"precision": 64.91, "recall": 67.83, "fscore": 66.34}, "DepthTrack_miss": {"precision": 57.57, "recall": 60.18, "fscore": 58.85}}};
+const LABEL={base_tuned:"Base model (tuned, best)",rung0:"rung-0 (seed 42)",rung0_seedB:"rung-0 (seed 123)",rung0_seedC:"rung-0 (seed 456)",
+  no_hallucinate:"no_hallucinate",no_recon_loss:"no_recon_loss",uni_a2r:"uni_a2r",uni_r2a:"uni_r2a",no_ortho:"no_ortho",
+  moe_small:"moe_small",moe_middle:"moe_middle",moe_hybrid:"moe_hybrid",
+  cma_fixed020:"cma_fixed020",no_distill:"no_distill",pmax_015:"pmax_015",pmax_025:"pmax_025",pmax_050:"pmax_050"};
+const SEEDS=["rung0","rung0_seedB","rung0_seedC"];
+const MECH=["no_hallucinate","no_recon_loss","uni_a2r","uni_r2a","no_ortho"];
+const MOE=["moe_small","moe_middle","moe_hybrid","cma_fixed020","no_distill","pmax_015","pmax_025","pmax_050"];
+const ORDER=["base_tuned"].concat(SEEDS,MECH,MOE);
+
+function gv(c,ds,k){ const v=DATA[c][ds]; return v? v[k] : null; }
+
+/* Base model, grid-search-tuned thresholds vs. the ablation study's shared
+   untuned default (rung-0). Tuned numbers come from the gs_*_FINAL.json
+   search logs (see ABLATIONS.md); "confirmed" means the search log ends
+   with an explicit convergence marker ("steps at floor, stopping search" /
+   GRIDSEARCH_DONE), not just "last value tried." */
+const TUNED = {
+  RGBT234:          {p:"mpr", s:"msr", pv:93.37, sv:70.02, status:"confirmed"},
+  RGBT234_miss:     {p:"mpr", s:"msr", pv:85.71, sv:62.80, status:"confirmed"},
+  LasHeR:           {p:"pr",  s:"sr",  pv:77.32, sv:61.96, status:"confirmed"},
+  LasHeR_miss:      {p:"pr",  s:"sr",  pv:null,  sv:null,  status:"none"},
+  VisEvent:         {p:"pr",  s:"auc", pv:87.92, sv:71.996,status:"unconfirmed"},
+  VisEvent_miss:    {p:"pr",  s:"auc", pv:null,  sv:null,  status:"none"},
+  DepthTrack:       {p:null,  s:"fscore", pv:null, sv:67.83, status:"confirmed"},
+  DepthTrack_miss:  {p:null,  s:"fscore", pv:null, sv:58.79, status:"confirmed"},
+};
+const TUNED_ORDER=["RGBT234","RGBT234_miss","LasHeR","LasHeR_miss","VisEvent","VisEvent_miss","DepthTrack","DepthTrack_miss"];
+const TUNED_NICE={RGBT234:"RGBT234",RGBT234_miss:"RGBT234 (miss)",LasHeR:"LasHeR",LasHeR_miss:"LasHeR (miss)",
+  VisEvent:"VisEvent",VisEvent_miss:"VisEvent (miss)",DepthTrack:"DepthTrack",DepthTrack_miss:"DepthTrack (miss)"};
+const TUNED_METRIC_NICE={RGBT234:"MPR / MSR",RGBT234_miss:"MPR / MSR",LasHeR:"PR / SR",LasHeR_miss:"PR / SR",
+  VisEvent:"PR / AUC",VisEvent_miss:"PR / AUC",DepthTrack:"F",DepthTrack_miss:"F"};
+
+(function(){
+  let head='<thead><tr><th></th><th>metric</th><th>base model, tuned</th><th>ablation default (rung-0)</th><th>gap</th><th>status</th></tr></thead>';
+  let body='<tbody>';
+  TUNED_ORDER.forEach(ds=>{
+    const t=TUNED[ds];
+    const rung0Full=DATA.rung0[ds];
+    const badge = t.status==="confirmed" ? '<span style="color:var(--good)">converged</span>'
+                : t.status==="unconfirmed" ? '<span style="color:var(--warn)">search cut short</span>'
+                : '<span style="color:var(--slate-2)">no search run</span>';
+    let tunedStr, defaultStr, gapStr;
+    if (t.status==="none") {
+      const dp = t.p ? rung0Full[t.p] : null, ds_ = rung0Full[t.s];
+      tunedStr = "—";
+      defaultStr = t.p ? `${dp.toFixed(2)} / ${ds_.toFixed(2)}` : ds_.toFixed(2);
+      gapStr = "—";
+    } else {
+      const dp = t.p ? rung0Full[t.p] : null, ds_ = rung0Full[t.s];
+      tunedStr = t.p ? `${t.pv.toFixed(2)} / ${t.sv.toFixed(2)}` : t.sv.toFixed(2);
+      defaultStr = t.p ? `${dp.toFixed(2)} / ${ds_.toFixed(2)}` : ds_.toFixed(2);
+      gapStr = `+${(t.sv-ds_).toFixed(2)}`;
+    }
+    body+=`<tr><td>${TUNED_NICE[ds]}</td><td class="fam">${TUNED_METRIC_NICE[ds]}</td><td>${tunedStr}</td><td>${defaultStr}</td><td>${gapStr}</td><td style="text-align:right">${badge}</td></tr>`;
+  });
+  body+='</tbody>';
+  document.getElementById('tunedtbl').innerHTML=head+body;
+})();
+
+/* table: each dataset group shows Precision-style + Success-style, for full and miss */
+(function(){
+  // [dataset key, group label, precision-field key, precision label, success-field key, success label]
+  const groups=[
+    ["RGBT234","RGBT234","mpr","MPR","msr","MSR"],
+    ["LasHeR","LasHeR","pr","PR","sr","SR"],
+    ["VisEvent","VisEvent","pr","PR","auc","AUC"],
+    ["DepthTrack","DepthTrack","precision","Pr","fscore","F"],
+  ];
+  let head='<thead><tr><th rowspan="3"></th>';
+  groups.forEach(g=>head+=`<th colspan="4" class="grp">${g[1]}</th>`);
+  head+='</tr><tr>';
+  groups.forEach(()=>head+='<th colspan="2">full</th><th colspan="2" class="miss">miss</th>');
+  head+='</tr><tr>';
+  groups.forEach(g=>head+=`<th>${g[3]}</th><th>${g[5]}</th><th class="miss">${g[3]}</th><th class="miss">${g[5]}</th>`);
+  head+='</tr></thead>';
+  let body='<tbody>';
+  ORDER.forEach((c,i)=>{
+    let cls=c.startsWith("rung0")?"base":"";
+    if(c==="rung0_seedC") cls+=" baseend";
+    if(c===MECH[0]||c===MOE[0]) cls+=" grpstart";
+    if(c==="base_tuned") cls+="tunedrow baseend";
+    body+=`<tr class="${cls}"><td>${LABEL[c]}${c==="rung0"?' <span class="fam">baseline</span>':''}${c==="base_tuned"?' <span class="fam">different, tuned thresholds — not apples-to-apples with rows below</span>':''}</td>`;
+    groups.forEach(([ds,,pk,,sk])=>{
+      const full=DATA[c][ds], miss=DATA[c][ds+"_miss"];
+      const fp=full&&full[pk]!=null?full[pk].toFixed(2):'-';
+      const fs=full&&full[sk]!=null?full[sk].toFixed(2):'-';
+      const mp=miss&&miss[pk]!=null?miss[pk].toFixed(2):'-';
+      const ms=miss&&miss[sk]!=null?miss[sk].toFixed(2):'-';
+      body+=`<td>${fp}</td><td>${fs}</td><td class="miss">${mp}</td><td class="miss">${ms}</td>`;
+    });
+    body+='</tr>';
+  });
+  body+='</tbody>';
+  document.getElementById('tbl').innerHTML=head+body;
+})();
+
+/* noise band chart */
+(function(){
+  const MISS=[["RGBT234_miss","msr","RGBT234 (MSR)"],["LasHeR_miss","sr","LasHeR (SR)"],
+    ["VisEvent_miss","auc","VisEvent (AUC)"],["DepthTrack_miss","fscore","DepthTrack (F)"]];
+  const ABL=MECH.concat(MOE);
+  let html='';
+  MISS.forEach(([d,m,nice])=>{
+    const seedVals=SEEDS.map(c=>gv(c,d,m));
+    const bMin=Math.min(...seedVals), bMax=Math.max(...seedVals);
+    const ablVals=ABL.map(c=>gv(c,d,m));
+    const all=seedVals.concat(ablVals);
+    let lo=Math.min(...all), hi=Math.max(...all); const pad=(hi-lo)*0.2+0.15; lo-=pad; hi+=pad;
+    const pos=v=>((v-lo)/(hi-lo))*100;
+    let outCount=0, dots='';
+    ABL.forEach((c,i)=>{
+      const v=ablVals[i]; const inside=v>=bMin-1e-9 && v<=bMax+1e-9; if(!inside)outCount++;
+      dots+=`<span class="dot ${inside?'in':'out'}" style="left:${pos(v)}%" title="${LABEL[c]}: ${v.toFixed(2)}"></span>`;
+    });
+    html+=`<div class="brow">
+      <div class="lbl">${nice}</div>
+      <div class="track"><div class="axis"></div>
+        <div class="bandbox" style="left:${pos(bMin)}%;width:${Math.max(pos(bMax)-pos(bMin),1)}%"></div>
+        ${dots}
+      </div>
+      <div class="verdict">${outCount}/${ABL.length} outside · band ${(bMax-bMin).toFixed(2)}</div>
+    </div>`;
+  });
+  document.getElementById('bands').innerHTML=html;
+})();
+</script>

ablation_results/official_metrics/mcitrack_b224_56_abl_cma_fixed020/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_cma_fixed020",
+  "RGBT234": {
+    "mpr": 91.7,
+    "msr": 69.18
+  },
+  "RGBT234_miss": {
+    "mpr": 84.16,
+    "msr": 62.53
+  },
+  "LasHeR": {
+    "pr": 76.27,
+    "sr": 61.3
+  },
+  "LasHeR_miss": {
+    "pr": 66.54,
+    "sr": 53.68
+  },
+  "VisEvent": {
+    "auc": 72.41,
+    "pr": 88.13
+  },
+  "VisEvent_miss": {
+    "auc": 61.45,
+    "pr": 77.02
+  },
+  "DepthTrack": {
+    "precision": 65.92,
+    "recall": 68.8,
+    "fscore": 67.33
+  },
+  "DepthTrack_miss": {
+    "precision": 56.19,
+    "recall": 58.64,
+    "fscore": 57.39
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_moe_hybrid/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_hybrid",
+  "RGBT234": {
+    "mpr": 92.12,
+    "msr": 69.75
+  },
+  "RGBT234_miss": {
+    "mpr": 84.52,
+    "msr": 63.07
+  },
+  "LasHeR": {
+    "pr": 76.2,
+    "sr": 61.07
+  },
+  "LasHeR_miss": {
+    "pr": 66.66,
+    "sr": 53.66
+  },
+  "VisEvent": {
+    "auc": 72.76,
+    "pr": 88.6
+  },
+  "VisEvent_miss": {
+    "auc": 61.3,
+    "pr": 76.87
+  },
+  "DepthTrack": {
+    "precision": 64.71,
+    "recall": 67.56,
+    "fscore": 66.11
+  },
+  "DepthTrack_miss": {
+    "precision": 56.24,
+    "recall": 58.81,
+    "fscore": 57.5
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_moe_middle/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_middle",
+  "RGBT234": {
+    "mpr": 91.9,
+    "msr": 69.14
+  },
+  "RGBT234_miss": {
+    "mpr": 85.81,
+    "msr": 63.44
+  },
+  "LasHeR": {
+    "pr": 77.02,
+    "sr": 61.74
+  },
+  "LasHeR_miss": {
+    "pr": 67.7,
+    "sr": 54.46
+  },
+  "VisEvent": {
+    "auc": 72.03,
+    "pr": 87.58
+  },
+  "VisEvent_miss": {
+    "auc": 61.19,
+    "pr": 76.59
+  },
+  "DepthTrack": {
+    "precision": 60.95,
+    "recall": 63.7,
+    "fscore": 62.3
+  },
+  "DepthTrack_miss": {
+    "precision": 54.73,
+    "recall": 57.22,
+    "fscore": 55.95
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_moe_small/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_small",
+  "RGBT234": {
+    "mpr": 91.9,
+    "msr": 69.42
+  },
+  "RGBT234_miss": {
+    "mpr": 85.23,
+    "msr": 63.61
+  },
+  "LasHeR": {
+    "pr": 75.63,
+    "sr": 60.52
+  },
+  "LasHeR_miss": {
+    "pr": 67.64,
+    "sr": 54.28
+  },
+  "VisEvent": {
+    "auc": 72.15,
+    "pr": 87.93
+  },
+  "VisEvent_miss": {
+    "auc": 61.24,
+    "pr": 76.7
+  },
+  "DepthTrack": {
+    "precision": 65.79,
+    "recall": 68.74,
+    "fscore": 67.23
+  },
+  "DepthTrack_miss": {
+    "precision": 54.38,
+    "recall": 56.77,
+    "fscore": 55.55
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_no_distill/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_no_distill",
+  "RGBT234": {
+    "mpr": 92.37,
+    "msr": 69.22
+  },
+  "RGBT234_miss": {
+    "mpr": 86.41,
+    "msr": 64.01
+  },
+  "LasHeR": {
+    "pr": 76.89,
+    "sr": 61.8
+  },
+  "LasHeR_miss": {
+    "pr": 66.94,
+    "sr": 53.91
+  },
+  "VisEvent": {
+    "auc": 71.97,
+    "pr": 87.93
+  },
+  "VisEvent_miss": {
+    "auc": 62.18,
+    "pr": 78.02
+  },
+  "DepthTrack": {
+    "precision": 64.66,
+    "recall": 67.5,
+    "fscore": 66.05
+  },
+  "DepthTrack_miss": {
+    "precision": 57.62,
+    "recall": 60.07,
+    "fscore": 58.82
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_no_hallucinate/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_no_hallucinate",
+  "RGBT234": {
+    "mpr": 92.41,
+    "msr": 69.45
+  },
+  "RGBT234_miss": {
+    "mpr": 84.57,
+    "msr": 62.68
+  },
+  "LasHeR": {
+    "pr": 76.14,
+    "sr": 61.03
+  },
+  "LasHeR_miss": {
+    "pr": 67.36,
+    "sr": 54.01
+  },
+  "VisEvent": {
+    "auc": 72.6,
+    "pr": 88.63
+  },
+  "VisEvent_miss": {
+    "auc": 61.54,
+    "pr": 77.18
+  },
+  "DepthTrack": {
+    "precision": 63.89,
+    "recall": 66.77,
+    "fscore": 65.3
+  },
+  "DepthTrack_miss": {
+    "precision": 56.11,
+    "recall": 58.64,
+    "fscore": 57.35
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_no_ortho/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_no_ortho",
+  "RGBT234": {
+    "mpr": 92.27,
+    "msr": 69.41
+  },
+  "RGBT234_miss": {
+    "mpr": 86.27,
+    "msr": 63.83
+  },
+  "LasHeR": {
+    "pr": 76.9,
+    "sr": 61.55
+  },
+  "LasHeR_miss": {
+    "pr": 65.94,
+    "sr": 53.09
+  },
+  "VisEvent": {
+    "auc": 72.36,
+    "pr": 88.42
+  },
+  "VisEvent_miss": {
+    "auc": 61.15,
+    "pr": 76.74
+  },
+  "DepthTrack": {
+    "precision": 64.24,
+    "recall": 67.06,
+    "fscore": 65.62
+  },
+  "DepthTrack_miss": {
+    "precision": 58.79,
+    "recall": 61.36,
+    "fscore": 60.05
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_no_recon_loss/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_no_recon_loss",
+  "RGBT234": {
+    "mpr": 91.57,
+    "msr": 68.94
+  },
+  "RGBT234_miss": {
+    "mpr": 85.2,
+    "msr": 62.98
+  },
+  "LasHeR": {
+    "pr": 76.28,
+    "sr": 61.16
+  },
+  "LasHeR_miss": {
+    "pr": 66.13,
+    "sr": 53.35
+  },
+  "VisEvent": {
+    "auc": 72.06,
+    "pr": 87.48
+  },
+  "VisEvent_miss": {
+    "auc": 61.52,
+    "pr": 77.2
+  },
+  "DepthTrack": {
+    "precision": 63.93,
+    "recall": 66.74,
+    "fscore": 65.3
+  },
+  "DepthTrack_miss": {
+    "precision": 54.81,
+    "recall": 57.27,
+    "fscore": 56.01
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_pmax_015/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_pmax_015",
+  "RGBT234": {
+    "mpr": 91.55,
+    "msr": 68.82
+  },
+  "RGBT234_miss": {
+    "mpr": 85.06,
+    "msr": 63.06
+  },
+  "LasHeR": {
+    "pr": 75.68,
+    "sr": 60.78
+  },
+  "LasHeR_miss": {
+    "pr": 65.83,
+    "sr": 53.1
+  },
+  "VisEvent": {
+    "auc": 72.59,
+    "pr": 88.29
+  },
+  "VisEvent_miss": {
+    "auc": 61.14,
+    "pr": 76.55
+  },
+  "DepthTrack": {
+    "precision": 62.33,
+    "recall": 65.18,
+    "fscore": 63.72
+  },
+  "DepthTrack_miss": {
+    "precision": 55.75,
+    "recall": 58.15,
+    "fscore": 56.92
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_pmax_025/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_pmax_025",
+  "RGBT234": {
+    "mpr": 92.23,
+    "msr": 69.14
+  },
+  "RGBT234_miss": {
+    "mpr": 85.13,
+    "msr": 63.1
+  },
+  "LasHeR": {
+    "pr": 76.35,
+    "sr": 61.2
+  },
+  "LasHeR_miss": {
+    "pr": 66.73,
+    "sr": 53.68
+  },
+  "VisEvent": {
+    "auc": 71.26,
+    "pr": 86.81
+  },
+  "VisEvent_miss": {
+    "auc": 62.41,
+    "pr": 78.51
+  },
+  "DepthTrack": {
+    "precision": 64.18,
+    "recall": 67.02,
+    "fscore": 65.57
+  },
+  "DepthTrack_miss": {
+    "precision": 53.71,
+    "recall": 56.16,
+    "fscore": 54.91
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_pmax_050/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_pmax_050",
+  "RGBT234": {
+    "mpr": 91.22,
+    "msr": 68.89
+  },
+  "RGBT234_miss": {
+    "mpr": 83.44,
+    "msr": 62.24
+  },
+  "LasHeR": {
+    "pr": 76.02,
+    "sr": 61.06
+  },
+  "LasHeR_miss": {
+    "pr": 66.49,
+    "sr": 53.51
+  },
+  "VisEvent": {
+    "auc": 72.33,
+    "pr": 88.05
+  },
+  "VisEvent_miss": {
+    "auc": 60.96,
+    "pr": 76.14
+  },
+  "DepthTrack": {
+    "precision": 64.91,
+    "recall": 67.83,
+    "fscore": 66.34
+  },
+  "DepthTrack_miss": {
+    "precision": 57.57,
+    "recall": 60.18,
+    "fscore": 58.85
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_rung0/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_rung0",
+  "RGBT234": {
+    "mpr": 91.94,
+    "msr": 69.13
+  },
+  "RGBT234_miss": {
+    "mpr": 84.55,
+    "msr": 62.66
+  },
+  "LasHeR": {
+    "pr": 76.33,
+    "sr": 61.29
+  },
+  "LasHeR_miss": {
+    "pr": 67.63,
+    "sr": 54.41
+  },
+  "VisEvent": {
+    "auc": 72.33,
+    "pr": 88.18
+  },
+  "VisEvent_miss": {
+    "auc": 62.2,
+    "pr": 78.06
+  },
+  "DepthTrack": {
+    "precision": 65.55,
+    "recall": 68.52,
+    "fscore": 67.01
+  },
+  "DepthTrack_miss": {
+    "precision": 56.61,
+    "recall": 59.21,
+    "fscore": 57.88
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_rung0_seedB/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_rung0_seedB",
+  "RGBT234": {
+    "mpr": 92.66,
+    "msr": 69.59
+  },
+  "RGBT234_miss": {
+    "mpr": 85.02,
+    "msr": 63.23
+  },
+  "LasHeR": {
+    "pr": 76.52,
+    "sr": 61.57
+  },
+  "LasHeR_miss": {
+    "pr": 67.37,
+    "sr": 54.28
+  },
+  "VisEvent": {
+    "auc": 72.28,
+    "pr": 88.11
+  },
+  "VisEvent_miss": {
+    "auc": 61.61,
+    "pr": 77.21
+  },
+  "DepthTrack": {
+    "precision": 63.86,
+    "recall": 66.61,
+    "fscore": 65.2
+  },
+  "DepthTrack_miss": {
+    "precision": 55.75,
+    "recall": 58.07,
+    "fscore": 56.89
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_rung0_seedC/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_rung0_seedC",
+  "RGBT234": {
+    "mpr": 91.52,
+    "msr": 69.29
+  },
+  "RGBT234_miss": {
+    "mpr": 84.69,
+    "msr": 62.89
+  },
+  "LasHeR": {
+    "pr": 76.67,
+    "sr": 61.37
+  },
+  "LasHeR_miss": {
+    "pr": 66.42,
+    "sr": 53.42
+  },
+  "VisEvent": {
+    "auc": 72.42,
+    "pr": 88.08
+  },
+  "VisEvent_miss": {
+    "auc": 61.61,
+    "pr": 77.04
+  },
+  "DepthTrack": {
+    "precision": 61.66,
+    "recall": 64.29,
+    "fscore": 62.95
+  },
+  "DepthTrack_miss": {
+    "precision": 56.58,
+    "recall": 59.14,
+    "fscore": 57.83
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_uni_a2r/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_uni_a2r",
+  "RGBT234": {
+    "mpr": 92.27,
+    "msr": 69.63
+  },
+  "RGBT234_miss": {
+    "mpr": 84.34,
+    "msr": 62.68
+  },
+  "LasHeR": {
+    "pr": 75.98,
+    "sr": 60.94
+  },
+  "LasHeR_miss": {
+    "pr": 66.81,
+    "sr": 53.84
+  },
+  "VisEvent": {
+    "auc": 71.66,
+    "pr": 87.23
+  },
+  "VisEvent_miss": {
+    "auc": 61.35,
+    "pr": 76.8
+  },
+  "DepthTrack": {
+    "precision": 65.13,
+    "recall": 67.99,
+    "fscore": 66.53
+  },
+  "DepthTrack_miss": {
+    "precision": 58.32,
+    "recall": 60.78,
+    "fscore": 59.53
+  }
+}

ablation_results/official_metrics/mcitrack_b224_56_abl_uni_r2a/metrics.json

@@ -0,0 +1,37 @@
+{
+  "config": "mcitrack_b224_56_abl_uni_r2a",
+  "RGBT234": {
+    "mpr": 92.23,
+    "msr": 69.15
+  },
+  "RGBT234_miss": {
+    "mpr": 83.93,
+    "msr": 62.31
+  },
+  "LasHeR": {
+    "pr": 76.03,
+    "sr": 60.89
+  },
+  "LasHeR_miss": {
+    "pr": 67.27,
+    "sr": 54.15
+  },
+  "VisEvent": {
+    "auc": 72.48,
+    "pr": 88.2
+  },
+  "VisEvent_miss": {
+    "auc": 61.35,
+    "pr": 76.78
+  },
+  "DepthTrack": {
+    "precision": 63.27,
+    "recall": 66.04,
+    "fscore": 64.62
+  },
+  "DepthTrack_miss": {
+    "precision": 57.53,
+    "recall": 60.0,
+    "fscore": 58.74
+  }
+}

ablation_results/raw_metrics/mcitrack_b224_56/metrics.json

@@ -0,0 +1,21 @@
+{
+  "config": "mcitrack_b224_56",
+  "RGBT234": {
+    "auc": 67.0,
+    "pr": 90.16
+  },
+  "RGBT234_miss": {
+    "auc": 60.26,
+    "pr": 83.8
+  },
+  "VisEvent": {
+    "auc": 62.55,
+    "pr": 75.53
+  },
+  "VisEvent_miss": {
+    "auc": 52.46,
+    "pr": 65.74
+  },
+  "DepthTrack": null,
+  "DepthTrack_miss": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_cma_fixed020/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_cma_fixed020",
+  "RGBT234": {
+    "auc": 67.18,
+    "pr": 90.06
+  },
+  "RGBT234_miss": {
+    "auc": 60.36,
+    "pr": 83.4
+  },
+  "LasHeR": {
+    "auc": 57.21,
+    "pr": 68.61
+  },
+  "LasHeR_miss": {
+    "auc": 51.47,
+    "pr": 60.65
+  },
+  "VisEvent": {
+    "auc": 63.78,
+    "pr": 76.76
+  },
+  "VisEvent_miss": {
+    "auc": 53.75,
+    "pr": 67.09
+  },
+  "DepthTrack": {
+    "precision": 65.92,
+    "recall": 68.8,
+    "fscore": 67.33
+  },
+  "DepthTrack_miss": {
+    "precision": 56.19,
+    "recall": 58.64,
+    "fscore": 57.39
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_big/mcitrack_b224_56_abl_moe_big_eval_RGBT234.log

@@ -0,0 +1,301 @@
+2026-07-06 17:16:11.648290: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:25.456600: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:25.759036: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:25.902932: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:25.982763: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.029982: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.035123: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.113810: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.290158: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.303732: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.399170: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.409458: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.472576: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.481117: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.554580: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.571015: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.582232: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.655245: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.667757: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.701321: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.769889: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.781423: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.794632: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.867149: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.903704: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:26.947503: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.098735: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.298994: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.315317: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.497491: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.583932: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.721471: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:27.797757: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+-1 afterrain
+-1 aftertree
+-1 baby
+-1 baginhand
+-1 baketballwaliking
+-1 balancebike
+-1 basketball2
+-1 bicyclecity
+-1 bike
+-1 bikeman
+-1 bikemove1
+-1 biketwo
+-1 blackwoman
+-1 blueCar
+-1 bluebike
+-1 boundaryandfast
+-1 bus6
+-1 call
+-1 car
+-1 car10
+-1 car20
+-1 car3
+-1 car37
+-1 car4
+-1 car41
+-1 car66
+-1 carLight
+-1 caraftertree
+-1 carnotfar
+-1 carnotmove
+-1 carred
+-1 child
+-1 child1
+-1 child3
+-1 child4
+-1 children2
+-1 children3
+-1 children4
+-1 crossroad
+-1 crouch
+-1 cycle1
+-1 cycle2
+-1 cycle3
+-1 cycle4
+-1 cycle5
+-1 diamond
+-1 dog
+-1 dog1
+-1 dog10
+-1 dog11
+-1 elecbike
+-1 elecbike10
+-1 elecbike2
+-1 elecbike3
+-1 elecbikechange2
+-1 elecbikeinfrontcar
+-1 elecbikewithhat
+-1 elecbikewithlight
+-1 elecbikewithlight1
+-1 face1
+-1 floor-1
+-1 flower1
+-1 flower2
+-1 fog
+-1 fog6
+-1 glass
+-1 glass2
+-1 graycar2
+-1 green
+-1 greentruck
+-1 greyman
+-1 greywoman
+-1 guidepost
+-1 hotglass
+-1 hotkettle
+-1 inglassandmobile
+-1 jump
+-1 kettle
+-1 kite2
+-1 kite4
+-1 luggage
+-1 man2
+-1 man22
+-1 man23
+-1 man24
+-1 man26
+-1 man28
+-1 man29
+-1 man3
+-1 man4
+-1 man45
+-1 man5
+-1 man55
+-1 man68
+-1 man69
+-1 man7
+-1 man8
+-1 man88
+-1 man9
+-1 manafterrain
+-1 mancross
+-1 mancross1
+-1 mancrossandup
+-1 mandrivecar
+-1 manfaraway
+-1 maninblack
+-1 maninglass
+-1 maningreen2
+-1 maninred
+-1 manlight
+-1 manoccpart
+-1 manonboundary
+-1 manonelecbike
+-1 manontricycle
+-1 manout2
+-1 manup
+-1 manwithbag
+-1 manwithbag4
+-1 manwithbasketball
+-1 manwithluggage
+-1 manwithumbrella
+-1 manypeople
+-1 manypeople1
+-1 manypeople2
+-1 mobile
+-1 night2
+-1 nightcar
+-1 nightrun
+-1 nightthreepeople
+-1 notmove
+-1 oldman
+-1 oldman2
+-1 oldwoman
+-1 orangeman1
+-1 people
+-1 people1
+-1 people3
+-1 playsoccer
+-1 push
+-1 rainingwaliking
+-1 raningcar
+-1 redbag
+-1 redcar
+-1 redcar2
+-1 redmanchange
+-1 rmo
+-1 run
+-1 run1
+-1 run2
+-1 scooter
+-1 shake
+-1 shoeslight
+-1 single1
+-1 single3
+-1 soccer
+-1 soccer2
+-1 soccerinhand
+-1 straw
+-1 stroller
+-1 supbus
+-1 supbus2
+-1 takeout
+-1 tallman
+-1 threeman
+-1 threeman2
+-1 threepeople
+-1 threewoman2
+-1 together
+-1 toy1
+-1 toy3
+-1 toy4
+-1 tree2
+-1 tree3
+-1 tree5
+-1 trees
+-1 tricycle
+-1 tricycle1
+-1 tricycle2
+-1 tricycle6
+-1 tricycle9
+-1 tricyclefaraway
+-1 tricycletwo
+-1 twoelecbike
+-1 twoelecbike1
+-1 twoman
+-1 twoman1
+-1 twoman2
+-1 twoperson
+-1 twowoman
+-1 twowoman1
+-1 walking40
+-1 walking41
+-1 walkingman
+-1 walkingman1
+-1 walkingman12
+-1 walkingman20
+-1 walkingman41
+-1 walkingmantiny
+-1 walkingnight
+-1 walkingtogether
+-1 walkingtogether1
+-1 walkingtogetherright
+-1 walkingwithbag1
+-1 walkingwithbag2
+-1 walkingwoman
+-1 whitebag
+-1 whitecar
+-1 whitecar3
+-1 whitecar4
+-1 whitecarafterrain
+-1 whiteman1
+-1 whitesuv
+-1 woamn46
+-1 woamnwithbike
+-1 woman
+-1 woman1
+-1 woman100
+-1 woman2
+-1 woman3
+-1 woman4
+-1 woman48
+-1 woman6
+-1 woman89
+-1 woman96
+-1 woman99
+-1 womancross
+-1 womanfaraway
+-1 womaninblackwithbike
+-1 womanleft
+-1 womanpink
+-1 womanred
+-1 womanrun
+-1 womanwithbag6
+-1 yellowcar
+Totally cost 13.8197660446167 seconds!

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_big/mcitrack_b224_56_abl_moe_big_eval_RGBT234_miss.log

@@ -0,0 +1,301 @@
+2026-07-06 17:16:38.393263: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:51.477357: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:51.590002: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:51.805504: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:51.926431: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.330615: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.489084: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.542198: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.622142: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.646215: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.657548: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.675722: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.833236: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.841257: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.870284: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:52.928307: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.047214: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.073083: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.076319: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.262914: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.317267: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.382204: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.550260: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.639602: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.734519: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.752908: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:53.780751: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:54.095929: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:54.382488: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:54.651375: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:55.050179: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:16:55.241823: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+-1 afterrain
+-1 aftertree
+-1 baby
+-1 baginhand
+-1 baketballwaliking
+-1 balancebike
+-1 basketball2
+-1 bicyclecity
+-1 bike
+-1 bikeman
+-1 bikemove1
+-1 biketwo
+-1 blackwoman
+-1 blueCar
+-1 bluebike
+-1 boundaryandfast
+-1 bus6
+-1 call
+-1 car
+-1 car10
+-1 car20
+-1 car3
+-1 car37
+-1 car4
+-1 car41
+-1 car66
+-1 carLight
+-1 caraftertree
+-1 carnotfar
+-1 carnotmove
+-1 carred
+-1 child
+-1 child1
+-1 child3
+-1 child4
+-1 children2
+-1 children3
+-1 children4
+-1 crossroad
+-1 crouch
+-1 cycle1
+-1 cycle2
+-1 cycle3
+-1 cycle4
+-1 cycle5
+-1 diamond
+-1 dog
+-1 dog1
+-1 dog10
+-1 dog11
+-1 elecbike
+-1 elecbike10
+-1 elecbike2
+-1 elecbike3
+-1 elecbikechange2
+-1 elecbikeinfrontcar
+-1 elecbikewithhat
+-1 elecbikewithlight
+-1 elecbikewithlight1
+-1 face1
+-1 floor-1
+-1 flower1
+-1 flower2
+-1 fog
+-1 fog6
+-1 glass
+-1 glass2
+-1 graycar2
+-1 green
+-1 greentruck
+-1 greyman
+-1 greywoman
+-1 guidepost
+-1 hotglass
+-1 hotkettle
+-1 inglassandmobile
+-1 jump
+-1 kettle
+-1 kite2
+-1 kite4
+-1 luggage
+-1 man2
+-1 man22
+-1 man23
+-1 man24
+-1 man26
+-1 man28
+-1 man29
+-1 man3
+-1 man4
+-1 man45
+-1 man5
+-1 man55
+-1 man68
+-1 man69
+-1 man7
+-1 man8
+-1 man88
+-1 man9
+-1 manafterrain
+-1 mancross
+-1 mancross1
+-1 mancrossandup
+-1 mandrivecar
+-1 manfaraway
+-1 maninblack
+-1 maninglass
+-1 maningreen2
+-1 maninred
+-1 manlight
+-1 manoccpart
+-1 manonboundary
+-1 manonelecbike
+-1 manontricycle
+-1 manout2
+-1 manup
+-1 manwithbag
+-1 manwithbag4
+-1 manwithbasketball
+-1 manwithluggage
+-1 manwithumbrella
+-1 manypeople
+-1 manypeople1
+-1 manypeople2
+-1 mobile
+-1 night2
+-1 nightcar
+-1 nightrun
+-1 nightthreepeople
+-1 notmove
+-1 oldman
+-1 oldman2
+-1 oldwoman
+-1 orangeman1
+-1 people
+-1 people1
+-1 people3
+-1 playsoccer
+-1 push
+-1 rainingwaliking
+-1 raningcar
+-1 redbag
+-1 redcar
+-1 redcar2
+-1 redmanchange
+-1 rmo
+-1 run
+-1 run1
+-1 run2
+-1 scooter
+-1 shake
+-1 shoeslight
+-1 single1
+-1 single3
+-1 soccer
+-1 soccer2
+-1 soccerinhand
+-1 straw
+-1 stroller
+-1 supbus
+-1 supbus2
+-1 takeout
+-1 tallman
+-1 threeman
+-1 threeman2
+-1 threepeople
+-1 threewoman2
+-1 together
+-1 toy1
+-1 toy3
+-1 toy4
+-1 tree2
+-1 tree3
+-1 tree5
+-1 trees
+-1 tricycle
+-1 tricycle1
+-1 tricycle2
+-1 tricycle6
+-1 tricycle9
+-1 tricyclefaraway
+-1 tricycletwo
+-1 twoelecbike
+-1 twoelecbike1
+-1 twoman
+-1 twoman1
+-1 twoman2
+-1 twoperson
+-1 twowoman
+-1 twowoman1
+-1 walking40
+-1 walking41
+-1 walkingman
+-1 walkingman1
+-1 walkingman12
+-1 walkingman20
+-1 walkingman41
+-1 walkingmantiny
+-1 walkingnight
+-1 walkingtogether
+-1 walkingtogether1
+-1 walkingtogetherright
+-1 walkingwithbag1
+-1 walkingwithbag2
+-1 walkingwoman
+-1 whitebag
+-1 whitecar
+-1 whitecar3
+-1 whitecar4
+-1 whitecarafterrain
+-1 whiteman1
+-1 whitesuv
+-1 woamn46
+-1 woamnwithbike
+-1 woman
+-1 woman1
+-1 woman100
+-1 woman2
+-1 woman3
+-1 woman4
+-1 woman48
+-1 woman6
+-1 woman89
+-1 woman96
+-1 woman99
+-1 womancross
+-1 womanfaraway
+-1 womaninblackwithbike
+-1 womanleft
+-1 womanpink
+-1 womanred
+-1 womanrun
+-1 womanwithbag6
+-1 yellowcar
+2026-07-06 17:16:55.506371: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+Totally cost 14.260838747024536 seconds!

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_big/mcitrack_b224_56_abl_moe_big_eval_depthtrack.log

@@ -0,0 +1,246 @@
+2026-07-06 17:23:19.058361: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:23.660729: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:23.797060: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:23.917363: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:23.951023: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:23.951024: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:23.989603: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.060638: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.110638: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.110667: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.131115: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.152701: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.240690: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.240698: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.257227: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.280150: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:24.303227: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:25.077223: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+-1 adapter01_indoor
+-1 backpack_indoor
+-1 bag01_indoor
+-1 bag02_indoor
+-1 ball01_wild
+-1 ball06_indoor
+-1 ball10_wild
+-1 ball11_wild
+-1 ball15_wild
+-1 ball18_indoor
+-1 ball20_indoor
+-1 bandlight_indoor
+-1 beautifullight02_indoor
+-1 book03_indoor
+-1 bottle04_indoor
+-1 card_indoor
+-1 cat01_indoor
+-1 colacan03_indoor
+-1 cube02_indoor
+-1 cube03_indoor
+-1 cube05_indoor
+-1 cup01_indoor
+-1 cup02_indoor
+-1 cup04_indoor
+-1 cup12_indoor
+-1 developmentboard_indoor
+-1 duck03_wild
+-1 dumbbells01_indoor
+-1 earphone01_indoor
+-1 file01_indoor
+-1 flag_indoor
+-1 glass01_indoor
+-1 hand01_indoor
+-1 human02_indoor
+-1 lock02_indoor
+-1 mobilephone03_indoor
+-1 notebook01_indoor
+-1 pigeon01_wild
+-1 pigeon02_wild
+-1 pigeon04_wild
+-1 pot_indoor
+-1 roller_indoor
+-1 shoes02_indoor
+-1 squirrel_wild
+-1 stick_indoor
+-1 toiletpaper01_indoor
+-1 toy02_indoor
+-1 toy09_indoor
+-1 ukulele01_indoor
+-1 yogurt_indoor
+Totally cost 4.920393705368042 seconds!
+Checking for new version
+A newer version of the VOT toolkit is available (0.8.1), please update.
+2 %
+4 %
+6 %
+8 %
+10 %
+12 %
+14 %
+16 %
+18 %
+20 %
+22 %
+24 %
+26 %
+28 %
+30 %
+32 %
+34 %
+36 %
+38 %
+40 %
+42 %
+44 %
+46 %
+48 %
+50 %
+52 %
+54 %
+56 %
+58 %
+60 %
+62 %
+64 %
+66 %
+68 %
+70 %
+72 %
+74 %
+76 %
+78 %
+80 %
+82 %
+84 %
+86 %
+88 %
+90 %
+92 %
+94 %
+96 %
+98 %
+100 %
+Loaded workspace in 'Depthtrack_workspace'
+Scanning registry /mnt/task_runtime/Depthtrack_workspace/trackers.ini
+Found 7 trackers
+Found data for 7 trackers
+Evaluating tracker mcitrack_b224_56_abl_moe_big_full
+2 %
+4 %
+6 %
+8 %
+10 %
+12 %
+14 %
+16 %
+18 %
+20 %
+22 %
+24 %
+26 %
+28 %
+30 %
+32 %
+34 %
+36 %
+38 %
+40 %
+42 %
+44 %
+46 %
+48 %
+50 %
+52 %
+54 %
+56 %
+58 %
+60 %
+62 %
+64 %
+66 %
+68 %
+70 %
+72 %
+74 %
+76 %
+78 %
+80 %
+82 %
+84 %
+86 %
+88 %
+90 %
+92 %
+94 %
+96 %
+98 %
+100 %
+Evaluating tracker mcitrack_b224_56_abl_moe_big_miss
+2 %
+4 %
+6 %
+8 %
+10 %
+12 %
+14 %
+16 %
+18 %
+20 %
+22 %
+24 %
+26 %
+28 %
+30 %
+32 %
+34 %
+36 %
+38 %
+40 %
+42 %
+44 %
+46 %
+48 %
+50 %
+52 %
+54 %
+56 %
+58 %
+60 %
+62 %
+64 %
+66 %
+68 %
+70 %
+72 %
+74 %
+76 %
+78 %
+80 %
+82 %
+84 %
+86 %
+88 %
+90 %
+92 %
+94 %
+96 %
+98 %
+100 %
+Evaluation concluded successfuly

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_big/mcitrack_b224_56_abl_moe_big_eval_depthtrack_miss.log

@@ -0,0 +1,97 @@
+2026-07-06 17:23:29.143986: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:33.946828: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.001415: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.029018: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.029036: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.054256: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.141032: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.227228: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.242690: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.253978: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.268415: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.291189: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.291189: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.330657: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:34.366747: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.106765: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.165229: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.175154: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.238619: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.252782: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.269037: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.370697: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+2026-07-06 17:23:35.398882: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
+To enable the following instructions: AVX2 AVX512F FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
+-1 adapter01_indoor
+-1 backpack_indoor
+-1 bag01_indoor
+-1 bag02_indoor
+-1 ball01_wild
+-1 ball06_indoor
+-1 ball10_wild
+-1 ball11_wild
+-1 ball15_wild
+-1 ball18_indoor
+-1 ball20_indoor
+-1 bandlight_indoor
+-1 beautifullight02_indoor
+-1 book03_indoor
+-1 bottle04_indoor
+-1 card_indoor
+-1 cat01_indoor
+-1 colacan03_indoor
+-1 cube02_indoor
+-1 cube03_indoor
+-1 cube05_indoor
+-1 cup01_indoor
+-1 cup02_indoor
+-1 cup04_indoor
+-1 cup12_indoor
+-1 developmentboard_indoor
+-1 duck03_wild
+-1 dumbbells01_indoor
+-1 earphone01_indoor
+-1 file01_indoor
+-1 flag_indoor
+-1 glass01_indoor
+-1 hand01_indoor
+-1 human02_indoor
+-1 lock02_indoor
+-1 mobilephone03_indoor
+-1 notebook01_indoor
+-1 pigeon01_wild
+-1 pigeon02_wild
+-1 pigeon04_wild
+-1 pot_indoor
+-1 roller_indoor
+-1 shoes02_indoor
+-1 squirrel_wild
+-1 stick_indoor
+-1 toiletpaper01_indoor
+-1 toy02_indoor
+-1 toy09_indoor
+-1 ukulele01_indoor
+-1 yogurt_indoor
+Totally cost 5.257458448410034 seconds!

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_big/metrics.json

@@ -0,0 +1,29 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_big",
+  "RGBT234": {
+    "auc": 67.25,
+    "pr": 90.07
+  },
+  "RGBT234_miss": {
+    "auc": 59.58,
+    "pr": 82.15
+  },
+  "VisEvent": {
+    "auc": 63.44,
+    "pr": 76.22
+  },
+  "VisEvent_miss": {
+    "auc": 53.48,
+    "pr": 66.74
+  },
+  "DepthTrack": {
+    "precision": 63.84,
+    "recall": 66.71,
+    "fscore": 65.25
+  },
+  "DepthTrack_miss": {
+    "precision": 56.49,
+    "recall": 59.02,
+    "fscore": 57.73
+  }
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_hybrid/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_hybrid",
+  "RGBT234": {
+    "auc": 67.2,
+    "pr": 89.59
+  },
+  "RGBT234_miss": {
+    "auc": 60.36,
+    "pr": 82.86
+  },
+  "LasHeR": {
+    "auc": 57.17,
+    "pr": 68.74
+  },
+  "LasHeR_miss": {
+    "auc": 50.5,
+    "pr": 59.75
+  },
+  "VisEvent": {
+    "auc": 64.11,
+    "pr": 77.17
+  },
+  "VisEvent_miss": {
+    "auc": 53.62,
+    "pr": 66.95
+  },
+  "DepthTrack": {
+    "precision": 64.71,
+    "recall": 67.56,
+    "fscore": 66.11
+  },
+  "DepthTrack_miss": {
+    "precision": 56.24,
+    "recall": 58.81,
+    "fscore": 57.5
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_middle/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_middle",
+  "RGBT234": {
+    "auc": 66.88,
+    "pr": 89.65
+  },
+  "RGBT234_miss": {
+    "auc": 60.41,
+    "pr": 83.52
+  },
+  "LasHeR": {
+    "auc": 57.62,
+    "pr": 69.54
+  },
+  "LasHeR_miss": {
+    "auc": 51.98,
+    "pr": 61.73
+  },
+  "VisEvent": {
+    "auc": 63.42,
+    "pr": 76.28
+  },
+  "VisEvent_miss": {
+    "auc": 53.51,
+    "pr": 66.71
+  },
+  "DepthTrack": {
+    "precision": 60.95,
+    "recall": 63.7,
+    "fscore": 62.3
+  },
+  "DepthTrack_miss": {
+    "precision": 54.73,
+    "recall": 57.22,
+    "fscore": 55.95
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_moe_small/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_moe_small",
+  "RGBT234": {
+    "auc": 66.55,
+    "pr": 89.0
+  },
+  "RGBT234_miss": {
+    "auc": 60.94,
+    "pr": 84.02
+  },
+  "LasHeR": {
+    "auc": 56.85,
+    "pr": 68.35
+  },
+  "LasHeR_miss": {
+    "auc": 51.76,
+    "pr": 61.35
+  },
+  "VisEvent": {
+    "auc": 63.54,
+    "pr": 76.59
+  },
+  "VisEvent_miss": {
+    "auc": 53.56,
+    "pr": 66.8
+  },
+  "DepthTrack": {
+    "precision": 65.79,
+    "recall": 68.74,
+    "fscore": 67.23
+  },
+  "DepthTrack_miss": {
+    "precision": 54.38,
+    "recall": 56.77,
+    "fscore": 55.55
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_no_distill/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_no_distill",
+  "RGBT234": {
+    "auc": 66.99,
+    "pr": 90.08
+  },
+  "RGBT234_miss": {
+    "auc": 60.99,
+    "pr": 84.22
+  },
+  "LasHeR": {
+    "auc": 57.8,
+    "pr": 69.2
+  },
+  "LasHeR_miss": {
+    "auc": 51.31,
+    "pr": 60.7
+  },
+  "VisEvent": {
+    "auc": 63.38,
+    "pr": 76.59
+  },
+  "VisEvent_miss": {
+    "auc": 54.43,
+    "pr": 67.96
+  },
+  "DepthTrack": {
+    "precision": 64.66,
+    "recall": 67.5,
+    "fscore": 66.05
+  },
+  "DepthTrack_miss": {
+    "precision": 57.62,
+    "recall": 60.07,
+    "fscore": 58.82
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_no_hallucinate/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_no_hallucinate",
+  "RGBT234": {
+    "auc": 66.87,
+    "pr": 89.81
+  },
+  "RGBT234_miss": {
+    "auc": 59.98,
+    "pr": 82.95
+  },
+  "LasHeR": {
+    "auc": 57.46,
+    "pr": 69.16
+  },
+  "LasHeR_miss": {
+    "auc": 51.86,
+    "pr": 61.33
+  },
+  "VisEvent": {
+    "auc": 63.97,
+    "pr": 77.2
+  },
+  "VisEvent_miss": {
+    "auc": 53.83,
+    "pr": 67.22
+  },
+  "DepthTrack": {
+    "precision": 63.89,
+    "recall": 66.77,
+    "fscore": 65.3
+  },
+  "DepthTrack_miss": {
+    "precision": 56.11,
+    "recall": 58.64,
+    "fscore": 57.35
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_no_ortho/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_no_ortho",
+  "RGBT234": {
+    "auc": 67.12,
+    "pr": 90.04
+  },
+  "RGBT234_miss": {
+    "auc": 60.3,
+    "pr": 83.48
+  },
+  "LasHeR": {
+    "auc": 57.17,
+    "pr": 68.72
+  },
+  "LasHeR_miss": {
+    "auc": 50.74,
+    "pr": 59.75
+  },
+  "VisEvent": {
+    "auc": 63.75,
+    "pr": 77.01
+  },
+  "VisEvent_miss": {
+    "auc": 53.48,
+    "pr": 66.84
+  },
+  "DepthTrack": {
+    "precision": 64.24,
+    "recall": 67.06,
+    "fscore": 65.62
+  },
+  "DepthTrack_miss": {
+    "precision": 58.79,
+    "recall": 61.36,
+    "fscore": 60.05
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_no_recon_loss/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_no_recon_loss",
+  "RGBT234": {
+    "auc": 66.25,
+    "pr": 89.01
+  },
+  "RGBT234_miss": {
+    "auc": 60.08,
+    "pr": 83.59
+  },
+  "LasHeR": {
+    "auc": 56.95,
+    "pr": 68.46
+  },
+  "LasHeR_miss": {
+    "auc": 51.0,
+    "pr": 60.01
+  },
+  "VisEvent": {
+    "auc": 63.45,
+    "pr": 76.2
+  },
+  "VisEvent_miss": {
+    "auc": 53.83,
+    "pr": 67.24
+  },
+  "DepthTrack": {
+    "precision": 63.93,
+    "recall": 66.74,
+    "fscore": 65.3
+  },
+  "DepthTrack_miss": {
+    "precision": 54.81,
+    "recall": 57.27,
+    "fscore": 56.01
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_pmax_015/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_pmax_015",
+  "RGBT234": {
+    "auc": 66.07,
+    "pr": 88.65
+  },
+  "RGBT234_miss": {
+    "auc": 60.3,
+    "pr": 83.38
+  },
+  "LasHeR": {
+    "auc": 57.34,
+    "pr": 69.06
+  },
+  "LasHeR_miss": {
+    "auc": 50.28,
+    "pr": 59.14
+  },
+  "VisEvent": {
+    "auc": 63.95,
+    "pr": 76.89
+  },
+  "VisEvent_miss": {
+    "auc": 53.48,
+    "pr": 66.67
+  },
+  "DepthTrack": {
+    "precision": 62.33,
+    "recall": 65.18,
+    "fscore": 63.72
+  },
+  "DepthTrack_miss": {
+    "precision": 55.75,
+    "recall": 58.15,
+    "fscore": 56.92
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_pmax_025/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_pmax_025",
+  "RGBT234": {
+    "auc": 66.81,
+    "pr": 89.78
+  },
+  "RGBT234_miss": {
+    "auc": 60.23,
+    "pr": 83.35
+  },
+  "LasHeR": {
+    "auc": 57.03,
+    "pr": 68.45
+  },
+  "LasHeR_miss": {
+    "auc": 51.1,
+    "pr": 60.33
+  },
+  "VisEvent": {
+    "auc": 62.71,
+    "pr": 75.61
+  },
+  "VisEvent_miss": {
+    "auc": 54.65,
+    "pr": 68.37
+  },
+  "DepthTrack": {
+    "precision": 64.18,
+    "recall": 67.02,
+    "fscore": 65.57
+  },
+  "DepthTrack_miss": {
+    "precision": 53.71,
+    "recall": 56.16,
+    "fscore": 54.91
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_pmax_050/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_pmax_050",
+  "RGBT234": {
+    "auc": 66.42,
+    "pr": 88.87
+  },
+  "RGBT234_miss": {
+    "auc": 59.47,
+    "pr": 81.69
+  },
+  "LasHeR": {
+    "auc": 57.53,
+    "pr": 69.08
+  },
+  "LasHeR_miss": {
+    "auc": 51.06,
+    "pr": 60.34
+  },
+  "VisEvent": {
+    "auc": 63.71,
+    "pr": 76.69
+  },
+  "VisEvent_miss": {
+    "auc": 53.32,
+    "pr": 66.31
+  },
+  "DepthTrack": {
+    "precision": 64.91,
+    "recall": 67.83,
+    "fscore": 66.34
+  },
+  "DepthTrack_miss": {
+    "precision": 57.57,
+    "recall": 60.18,
+    "fscore": 58.85
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_rung0/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_rung0",
+  "RGBT234": {
+    "auc": 66.34,
+    "pr": 89.09
+  },
+  "RGBT234_miss": {
+    "auc": 59.86,
+    "pr": 82.86
+  },
+  "LasHeR": {
+    "auc": 55.27,
+    "pr": 66.16
+  },
+  "LasHeR_miss": {
+    "auc": 50.59,
+    "pr": 59.55
+  },
+  "VisEvent": {
+    "auc": 63.7,
+    "pr": 76.81
+  },
+  "VisEvent_miss": {
+    "auc": 54.45,
+    "pr": 67.99
+  },
+  "DepthTrack": {
+    "precision": 65.55,
+    "recall": 68.52,
+    "fscore": 67.01
+  },
+  "DepthTrack_miss": {
+    "precision": 56.61,
+    "recall": 59.21,
+    "fscore": 57.88
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_rung0_seedB/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_rung0_seedB",
+  "RGBT234": {
+    "auc": 67.27,
+    "pr": 90.46
+  },
+  "RGBT234_miss": {
+    "auc": 60.18,
+    "pr": 82.63
+  },
+  "LasHeR": {
+    "auc": 56.0,
+    "pr": 67.17
+  },
+  "LasHeR_miss": {
+    "auc": 51.1,
+    "pr": 60.71
+  },
+  "VisEvent": {
+    "auc": 63.66,
+    "pr": 76.75
+  },
+  "VisEvent_miss": {
+    "auc": 53.91,
+    "pr": 67.25
+  },
+  "DepthTrack": {
+    "precision": 63.86,
+    "recall": 66.61,
+    "fscore": 65.2
+  },
+  "DepthTrack_miss": {
+    "precision": 55.75,
+    "recall": 58.07,
+    "fscore": 56.89
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_rung0_seedC/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_rung0_seedC",
+  "RGBT234": {
+    "auc": 66.73,
+    "pr": 89.04
+  },
+  "RGBT234_miss": {
+    "auc": 60.29,
+    "pr": 83.09
+  },
+  "LasHeR": {
+    "auc": 57.19,
+    "pr": 68.5
+  },
+  "LasHeR_miss": {
+    "auc": 50.83,
+    "pr": 59.93
+  },
+  "VisEvent": {
+    "auc": 63.8,
+    "pr": 76.72
+  },
+  "VisEvent_miss": {
+    "auc": 53.9,
+    "pr": 67.1
+  },
+  "DepthTrack": {
+    "precision": 61.66,
+    "recall": 64.29,
+    "fscore": 62.95
+  },
+  "DepthTrack_miss": {
+    "precision": 56.58,
+    "recall": 59.14,
+    "fscore": 57.83
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_uni_a2r/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_uni_a2r",
+  "RGBT234": {
+    "auc": 66.63,
+    "pr": 89.29
+  },
+  "RGBT234_miss": {
+    "auc": 59.86,
+    "pr": 82.5
+  },
+  "LasHeR": {
+    "auc": 56.81,
+    "pr": 68.24
+  },
+  "LasHeR_miss": {
+    "auc": 50.57,
+    "pr": 59.6
+  },
+  "VisEvent": {
+    "auc": 63.1,
+    "pr": 75.97
+  },
+  "VisEvent_miss": {
+    "auc": 53.66,
+    "pr": 66.89
+  },
+  "DepthTrack": {
+    "precision": 65.13,
+    "recall": 67.99,
+    "fscore": 66.53
+  },
+  "DepthTrack_miss": {
+    "precision": 58.32,
+    "recall": 60.78,
+    "fscore": 59.53
+  },
+  "VOT22RGBD": null
+}

ablation_results/raw_metrics/mcitrack_b224_56_abl_uni_r2a/metrics.json

@@ -0,0 +1,38 @@
+{
+  "config": "mcitrack_b224_56_abl_uni_r2a",
+  "RGBT234": {
+    "auc": 66.68,
+    "pr": 89.64
+  },
+  "RGBT234_miss": {
+    "auc": 59.57,
+    "pr": 82.23
+  },
+  "LasHeR": {
+    "auc": 57.38,
+    "pr": 69.0
+  },
+  "LasHeR_miss": {
+    "auc": 51.21,
+    "pr": 60.38
+  },
+  "VisEvent": {
+    "auc": 63.85,
+    "pr": 76.82
+  },
+  "VisEvent_miss": {
+    "auc": 53.67,
+    "pr": 66.87
+  },
+  "DepthTrack": {
+    "precision": 63.27,
+    "recall": 66.04,
+    "fscore": 64.62
+  },
+  "DepthTrack_miss": {
+    "precision": 57.53,
+    "recall": 60.0,
+    "fscore": 58.74
+  },
+  "VOT22RGBD": null
+}