3dsem

Classify point clouds with pretrained 3D semantic segmentation models, on your own machine, from one command.

pip install 3dsem

sem install dales-utonia
sem infer dales-utonia tile.las

The classified .laz appears next to your input file, with per-point classification, confidence, and every original dimension carried over.

What you need

• An NVIDIA GPU with a current driver (Windows 527.41+, Linux 525.60.13+)
• Python 3.9 or newer
• About 10 GB of disk per model

sem install shows the download size, the license, and a GPU check, then asks once before any bytes move. It is safe to interrupt and resumes where it stopped. After install, inference runs fully offline.

Models

dales-utonia is trained on DALES aerial LiDAR and predicts 7 classes: ground, vegetation, vehicle, powerline, fence, pole, building.

dales-hag-utonia predicts the same 7 classes from the same data, with height above ground added as an input. It scores higher on held-out DALES scenes (0.86 mIoU against 0.98 overall accuracy) and is the one to reach for first.

ieee-utonia is for low density airborne LiDAR and predicts 5 classes: ground, vegetation, building, water, bridge deck. Alongside geometry it reads return number, which a LAS carries under that name, so it runs with no extra flags.

h3d-utonia is for coloured photogrammetric clouds and predicts 8 classes: ground, vehicle, urban furniture, roof, wall, low veg, tree, chimneys. It is the one model that consumes colour, so it needs --rgb-max naming the full-scale value your file uses:

sem infer h3d-utonia tile.laz --rgb-max 255

Every model is licensed CC-BY-NC 4.0 (non-commercial), shown before install, and needs about 3.5 GB of download.

sem models lists what is available and installed. Bare sem opens an interactive picker.

How do I make it more accurate?

Use a preset. Each one turns on more of the same three ideas: predict the scene from several augmented views and vote (test-time augmentation), tile the scene a second time at a half offset so no point sits on a tile edge (overlapped voting), and clean up the labels afterwards (smoothing, island removal, geometry rules).

• default: one pass with light smoothing
• --med: 4 voting views, island removal, single tiling pass. Roughly 4x.
• --high: 6 views including flips and rotations, stronger smoothing, and the overlapped second pass. Roughly 24x, because the views and the second pass multiply.
• --ultra: 9 views, the strongest smoothing, overlapped. Roughly 36x.

Geometry rules are never part of a preset. They rewrite labels, so they need --rules with --roles.

An explicit flag always wins over a preset, so --ultra --no-sieve means ultra without island removal.

How accurate is it on my data?

If you have a tile whose classification you trust, classify a copy and compare the two:

sem eval tile_pred.laz tile_truth.laz

No inference runs; it reads the Classification column of both files and prints, per class, IoU, precision, recall and F1 with point counts, then overall accuracy, mIoU, macro F1, Cohen's kappa, and a confusion matrix. Files written by sem infer carry their own class legend, whatever codes that run actually used, so rows read 2 ground rather than bare numbers. A file without a legend shows plain codes, and --names 1=ground,4=water names any code by hand. The files must be the same points in the same order, which is what sem infer's own export gives you. --pred-field and --truth-field name a different column on either side.

When the truth file uses different codes than the model, --map translates them first, and --ignore keeps codes out of the scoring:

sem eval tile_pred.laz tile_truth.laz --map 3=5,4=5 --ignore 0,1

folds the truth's low and medium vegetation into the model's single vegetation class and skips points that were never classified. --map 3,4=5 says the same thing: several truth codes go on the left of the =, one predicted code on the right.

To see whether the cleanup passes help or hurt on your data, run the inference with --keep-raw: it also exports a copy from before post-processing into a raw/ folder inside the job, under the same names. Score each against the same truth and compare:

sem eval tile_dales-hag-utonia_predictions/ truth/
sem eval tile_dales-hag-utonia_predictions/raw/ truth/

A whole batch works the same way with two folders: files pair by name (tile_pred.laz matches tile.laz), each tile gets a summary line, and the totals pool every point before computing IoU, so a big tile counts for more than a small one. A predicted file with no truth partner stops the run and is named. Folders pair with folders and files with files, never mixed.

How do I make it faster?

--low is a single pass with no cleanup, the fastest option. If a preset is mostly what you want, --no-overlap drops its second tiling pass, which is about half its extra cost.

Big buildings come out patchy or cut through. Why?

The scene is processed in square tiles, typically 50 m on a side. An object bigger than one tile is predicted in pieces, and the pieces can disagree. Two fixes that combine well:

• --chunk-xy 100 makes the tiles bigger, so a large building fits in one. Costs GPU memory.
• --overlap (on automatically with --high and up) predicts a second pass at a half offset and votes, which removes most seam artifacts. It costs about 4x on its own, so --med leaves it off.

It ran out of GPU memory

Lower --chunk-xy, try 35 and then 25. Smaller tiles need less VRAM, and the extra seams they create are what --overlap is for.

Poles or powerlines are disappearing

Presets from --med up turn on island removal, which absorbs clusters smaller than 10 points into their surroundings. Thin objects are exactly small clusters. Keep the filter but make it gentler with --sieve-min-pts 5, or turn it off with --no-sieve.

Can I hide the model's low-confidence guesses?

--unclass 0.6 exports every point below 60% confidence as unclassified instead of its best guess. Bare --unclass uses 0.5. To see where the model is unsure, --extra-dims adds per-point diagnostic fields: entropy (0..1 uncertainty), margin (gap between the top two classes), probs (one probability field per class, grows the file), and ood (the raw scores behind the gates below). Combine them freely: --extra-dims entropy,margin.

It confidently labels things it has never seen

A crane or a boat has no class to land in, so the model puts it in the nearest one it knows, often with high confidence. Confidence alone will not catch that, because the model is confident and wrong. Two other scores will.

--unclass-gmm compares each point against how the training classes actually looked to the model and unclassifies anything too far from all of them. Bare, it uses the threshold the model was shipped with; give it a number to override. --unclass-maxlogit 4.0 catches the opposite case, points where no class drew much evidence at all. They find different mistakes, so using both is normal.

To choose your own numbers, run once with --extra-dims ood, which writes the raw scores into the output file so you can see where your data sits, then set the thresholds and re-export. Re-exporting never re-runs the model.

Every model in one folder ran, but the gates did nothing

--unclass-gmm needs a model that shipped with those statistics. If a model predates them, that flag stops with an error naming what is missing rather than silently exporting ungated results. --unclass and --unclass-maxlogit work on any model.

Part of my file is already classified correctly

--input-field preserve:Classification:2,9 keeps the codes your file already carries for those points and writes the model's prediction everywhere else. They are read from the source file at export, so inference, the unclassified gates and the cleanup passes never touch them.

Kept codes are written as your file spells them. When a kept code means something different in the model's output (your water 9 has no model class), add --preserve-map 9=2 to write it as ground instead, and the export warns whenever a kept code is not one the model uses, so a foreign code never slips through silently.

The export reports how often the model disagreed with what you preserved, which is a quick check on both.

You can also do it afterwards, on a file that is already classified:

sem preserve tile_pred.laz tile.las --keep Classification:2,9

That writes tile_pred_preserved.laz beside it. Same overlay, decided after you have seen the result rather than before, with the same --map 9=2 option for codes that need translating. The two files have to be the same points in the same order, which is checked on counts and coordinates. The predicted file needs to be on ASPRS codes for this: the codes you keep come from your own file's scheme, so overlaying them onto raw model indices would put two vocabularies in one column. Export it with --asprs, which preserving during inference turns on for you.

Can I combine models?

sem infer dales-utonia+dales-hag-utonia tile.las runs every model in the chain and merges their predictions with a vote: each model's per-class probabilities are averaged, and the strongest combined evidence wins. Where the models agree, the label sticks; exact ties go to the model you listed first, so lead with your strongest. The result carries an agreement field (what fraction of models agreed on each point) and an ens_member field (which model drove each label). The models must share the same class list. Each member also keeps its own tile_<model>_predictions folder, so you can compare them individually.

Can I run my own trained model?

If a training run left you a folder holding final_model.pth and run.json, register it:

sem local-model C:\runs\my_run --name myrun

It appears in sem models and runs like any other: sem infer myrun tile.las. Nothing is copied; the folder is used in place, and sem local-model --forget myrun unregisters it without touching the files.

The weights are yours but the environment is not: inference borrows the installed env of a catalog model with the same backbone, so install one that matches first. Everything else the run needs (grid, features, classes) is read from the run.json beside the weights.

My file has no coordinate system

It still runs. sem uses the coordinates exactly as they stand, reprojects nothing, and warns that it is taking their unit on trust: everything below (tile size, voxel grid, neighbourhoods) is a length in metres, so a file already in metres is correct and a file in feet is off by that factor. Most aerial LiDAR is already metre-projected, which is why this usually just works.

--epsg 26917 declares the projection when you know it (use your own zone's code), which silences the warning and makes the exported cloud georeferenced.

The one case that stops is lon/lat-shaped coordinates, where the whole scene spans a fraction of a degree. A 50 m tile there would cover everything at once, so that is not a slightly worse answer, it is no answer; declare a projected CRS with --epsg and it proceeds.

Reprojection only ever happens to get your data into metres, from degrees or from feet. A model is not tied to any particular CRS, only to the scale it was trained at, so a cloud already in a metre projection is never transformed.

My file is not a LAS. How does sem know what its columns mean?

It does not, and it will not guess. LAS and LAZ name their dimensions in the format spec, so intensity, return number and classification are read straight off a LAS with no help from you. Every other format (.txt, .csv, .xyz, .pts, .npy, .npz, .ply, .pcd) leaves the meaning of a field entirely up to whoever wrote it, so sem asks you:

sem infer dales-utonia scan.txt \
    --input-field xyz:1,2,3 --input-field intensity:5

--input-field KEY:COLUMN[:VALUE] says where one thing lives in your file, repeated per entry. Columns are names or 0-based numbers.

• xyz:A,B,C and rgb:R,G,B take three columns
• intensity:C, return-number:C and hag:C take one
• ground:C:V,... and preserve:C:V,... add the value(s) in that column that mean it, because which column and which value in it are one idea

A model that wants a channel you did not name stops and lists the fields your file actually has.

Nothing is inferred from a column's position or from a name that looks familiar, because a file with id,x,y,z in that order and a file with x,y,z,id are indistinguishable to anything except you.

My colours come out black, or sem asks for --rgb-max

Point clouds store colour as 8, 10, 12 or 16 bit, and no format records which. Guessing it from the brightest point in the scene turns a 12-bit cloud nearly black and a dark 16-bit cloud into a blown-out one, so --rgb-max states the full-scale value: 255, 1023, 4095, 65535, or 1 for float 0 to 1 colour. It is only needed when the model actually consumes colour; a model that runs on intensity ignores the colour in your file and never asks. Of the shipped models that is h3d-utonia alone, which stops and asks for it rather than guessing a depth and returning a near-black or blown-out cloud.

Height above ground

Models trained with a HAG channel compute one at staging, using a ground raster whose cell size comes from the linear unit your file's CRS declares: 2 metres, or whatever length equals 2 metres in your file's own unit. A cloud in US survey feet gets a 6.56 foot cell, which is the same ground resolution. Nothing is assumed about your units; they are read from the CRS.

--hag-cell overrides that when you want a different resolution. The height error it costs is roughly the cell size times the local slope, so a smaller cell buys accuracy on steep ground and costs memory. A cloud that declares no projected CRS has no unit to read, so it stops until you either declare one with --epsg or state the cell with --hag-cell.

--ground-method picks where the ground comes from (smrf, csf, zmin, or labels when your file already marks ground). --csf-rigidness (1 steep, 2 moderate, 3 flat/urban), --smrf-window and --smrf-cut tune the ground filters. If your file marks ground with a class code, say so in one entry: --input-field ground:Classification:2. The column alone is not enough, because it does not say which value in it means ground. More than one code can count as ground: ground:Classification:2,9 also folds water under the ground surface, which keeps HAG flat across lakes and rivers.

The cleanup rules are wrong for my data

The geometry rules only run when you ask for them and tell them what your classes mean: --rules --roles veg=vegetation,building=building. sem no longer decides that a class is vegetation because its name contains "tree". Their thresholds are lengths in your scene's vertical unit and unitless ratios, all settable through --rule-set: ground_hag (0.1), lowveg_hag (0.35), high_hag (2.0), planar_min (0.55), scatter_min (0.4). For example --rule-set high_hag=3,planar_min=0.6.

Controlling the exported class codes

By default each class exports as its own model index: 0, 1, 2 and so on, in the order the model lists them. On the shipped DALES models that is ground 0, vegetation 1, vehicle 2, powerline 3, fence 4, pole 5, building 6, and gated points 7.

Those indices are not ASPRS codes. In the spec 2 is Ground, so a viewer reading your file per the spec sees the vehicles as ground. Exported las and laz carry a legend recording what each code meant, and sem eval reads it, but nothing outside sem does. Any file leaving sem wants the next flag.

--asprs exports spec codes instead: ground 2, vegetation 5, building 6, powerline 14, pole 15, and user-definable codes from 64 for classes the spec has no name for, which on these models is vehicle 64 and fence 65. A class the table does not name takes the next free code at or above 64.

--asprs-map ground=2,building=6 overrides any of them by name and implies --asprs. Two classes landing on one code stops the run rather than merging them silently. Preserving codes from your own file implies --asprs too, since those codes already speak it.

--unclassified-code sets the code gated points receive. It defaults to the class count, or to 1 under --asprs, which is ASPRS "Unclassified" and no class maps onto. Export stops rather than letting that code collide with a real class.

I don't have a GPU

sem classifies on your own machine and needs an NVIDIA GPU. There is no cloud option here. Conversion and export are CPU work, but the model itself is not.

Can I rerun with different settings without reconverting?

Pass --keep-npz on the first run, then yes.

The output folder is a self-contained job, named after your file and the model (tile_dales-utonia_predictions/). The staged .npz inside it hold the converted channels, and they are what a re-run reuses. By default they are deleted once the export is written, because most runs only want the classified file. --keep-npz retains them.

With them kept: quality options (presets, TTA, cleanup, export) never reconvert. Conversion options (--epsg, --ground-method, --hag, ...) and changes to the input file itself reconvert automatically; identical settings reuse the staged files. Every model keeps its own job folder, so switching models never mixes results.

Can I just name the output file?

Yes. An output ending in .laz, .las, .txt, .csv or .ply is the file itself, and its extension picks the format, so --format becomes redundant:

sem infer dales-utonia tile.las out/pc.laz

The job folder is created beside it (out/pc_dales-utonia_predictions/), so everything the run makes stays where you pointed. This is for a single input file; a folder of tiles produces one result per scene, so give it a folder.

If you have set a default with sem output, a bare filename lands there: sem infer dales-utonia tile.las pc.laz writes <your dir>/pc.laz. Put a folder in the path and it goes exactly there instead.

A whole folder of tiles?

Pass the folder. Every .las/.laz/.ply/.pcd inside becomes one job, and results land in <model>_predictions/ inside it.

Where does everything live?

Downloads go to ~/.trainer (set TRAINER_HOME to move them). Your data and results never go there: each job is a folder next to your input, or under a default you set with sem output <dir> and undo with sem output off. sem clean dales-utonia removes one model; sem clean --all removes everything sem ever downloaded.

Can I look at the intermediate files?

Run with --keep-npz, or the intermediates are gone once the export is written. Then sem tolaz job/tile_input.npz writes tile_input.laz beside it. The xyz becomes the cloud, rgb the color, and every other per-point channel a named field you can shade by in CloudCompare. That works on the staged input as well as on predictions, so it is how you see the features a model actually received, not just what it predicted.

Every option

sem infer --help documents all of it. Add --pick to browse and edit every option with arrow keys before running.

Licensing

The sem tool is MIT licensed. Each model ships a NOTICE.md stating its architecture credits and license terms; some models carry a non-commercial restriction inherited from their pretrained components, shown before you install.