Skip to content

README.md

Latest commit

 

History

History
73 lines (58 loc) · 3.58 KB

README.md

File metadata and controls

73 lines (58 loc) · 3.58 KB
Raw

DL-VINS TensorRT engine zoo

Serialized TensorRT engines for the deep front-end (feature extractors + LightGlue matchers). .engine files are gitignored and built per machine — a TensorRT engine is not portable across GPU architectures / TensorRT versions, so each machine (x86-64 dev box, Jetson, …) rebuilds its own engines into this tree. Only one platform's engines live in a checkout at a time;

Layout

weights/
├── matcher/                      # shared across ALL datasets (resolution-independent)
│   ├── mono_extractor/           # single-pair engines  (batch 1)
│   └── stereo_extractor/         # batched engines      (batch 2)
└── <dataset>/                    # euroc, ntu_viral, botanic_garden, subt_mrs, …
    ├── mono_extractor/           # extractor engines, batch 1
    └── stereo_extractor/         # extractor engines, batch 2
  • <dataset>/<role>/ holds extractor engines only, sized to that dataset's image resolution. <role> is mono_extractor (batch 1) or stereo_extractor (batch 2).
  • matcher/<role>/ holds matcher engines. These are resolution-independent (the engine takes (1, K, 2) keypoints + (1, K, D) descriptors; image size only enters as a runtime normalization scalar), so one shared set serves every dataset. The mono variant is single-pair (batch 1); the stereo variant is batched (_b2, for the temporal-L / stereo-LR passes fused into one call).

Weights layout

Filename convention

Engine metadata is encoded in the filename, not in extra folder levels:

<model>_b{N}_h{H}_w{W}_kp{K}_sim.fp16.engine        # extractor
<model>_lightglue[_b{N}]_kp{K}_sim.fp16.engine      # matcher
Field Meaning Example
<model> superpoint / aliked / raco / xfeat superpoint
_b{N} batch: 1 mono extractor, 2 stereo extractor, 2 stereo matcher _b2
_h{H}_w{W} input resolution (must be model-divisible; ALIKED ÷32) _h480_w752
_kp{K} max keypoint budget _kp256
_sim onnx-simplified
.fp16 precision

ALIKED is split into two extractor engines: a dense backbone aliked_n16_dense_b{N}_h{H}_w{W}_sim.fp16.engine (no _kp, outputs dense maps) and a descriptor head aliked_n16_dhead_b{N}_h{H}_w{W}_kp{K}_sim.fp16.engine. RaCo has no native descriptors and reuses the ALIKED descriptor head when matching.

Examples:

euroc/mono_extractor/superpoint_b1_h480_w752_kp256_sim.fp16.engine
botanic_garden/stereo_extractor/aliked_n16_dense_b2_h608_w960_sim.fp16.engine
matcher/mono_extractor/aliked_lightglue_kp256_sim.fp16.engine
matcher/stereo_extractor/superpoint_lightglue_b2_kp256_sim.fp16.engine

How engines are resolved at runtime

weights_folder (set in the dataset config / launch file) points at weights/<dataset>. FeatureTracker::resolveEnginePath() then:

  • Extractors: scan weights/<dataset>/<role>/ for an .engine whose name contains the required tags (model name, and _kp{K} when relevant).
  • Matchers: scan the sibling shared weights/matcher/<role>/ for an .engine containing the matcher type (lightglue), the extractor model name, and _kp{K}.

Selection is by substring match, so keep one engine per (model, role, kp) — duplicates trigger a warning and a non-deterministic pick.