Semantic Segmentation Assistant (Street View)

What you can do

Everything below happens in the app window. You do not need a second program to finish a figure.

• Import: open or drop one photo, or pick a folder for batch work (subfolders included). jpg, png, bmp, webp and tif work, including Chinese paths.
• Three scenes: Cityscapes (19 road classes), Mapillary (65 finer street classes: lane lines, barriers, manholes), ADE20K (about 150 indoor–outdoor classes).
• Models: SegFormer B0 is the default and runs on a normal laptop. Stronger machines can use B2; an NVIDIA GPU unlocks B5 or Mask2Former. No GPU means the app stays on CPU.
• Preview: switch Compare / Mask / Labels / Original, then drag mask opacity.
• Label map: drawn on the original photo, with class boundaries and names. Type size follows the image, so large photos stay readable without zooming in.
• Label language: Chinese or English. Use English for journal figures. The preview bar and the export window both have the switch.
• Export: colour-block map, mask overlay, label map and a class table. A class-id image is optional. Batch jobs write one combined CSV.
• Optional metrics: green-view, enclosure, sky openness and similar. They stay off until you tick them; click “?” for the formula.
• Original files stay untouched. Any shrinking happens in memory only.
• Single image for exploring; batch for a folder, with pause, continue and resume from the last finished photo.

Install and open

Windows: run StreetSegmentation-Setup-*.exe, then open Semantic Segmentation Assistant (Street View) from the Start menu. You do not need Python, CUDA, Visual C++ or any developer toolkit. If an NVIDIA GPU and driver are available the app uses the GPU automatically; otherwise it uses the CPU. If the GPU fails later it continues on CPU instead of quitting. There is no separate GPU installer. If antivirus flags the installer, add it to trusted files and try again.

macOS: open the .dmg and drag Semantic Segmentation Assistant (Street View) into Applications. If Gatekeeper says the developer cannot be verified, open System Settings → Privacy & Security and choose Open Anyway. Prefer Apple-notarised builds for machines that are not yours. Run it from Applications rather than from the disk image. Double-click still opens the window. The same CLI as Linux is available from Terminal: cd into the app’s Contents/MacOS folder and run ./StreetSegmentation.

Linux: there is no graphical window. You receive a .tar.gz. Unpack it and keep the whole StreetSegmentation folder — the program plus _internal. Do not copy the binary alone. In a terminal: tar -xzf 语义分割助手-街景版-linux-*.tar.gz, then cd StreetSegmentation, then activate online with the same code as Windows/macOS: ./StreetSegmentation activate YOUR-CODE. Headless servers work; the program does not need a display.

• No dedicated GPU: stay on SegFormer B0 and avoid very large images.
• NVIDIA GPU: B2, B5 or Mask2Former become practical.
• Apple silicon: the app can use the Apple GPU; large models are still slower.
• Under 8 GB of RAM: avoid B5 and the Mapillary model, and process fewer images at a time.

First-time activation

Paste the activation code you were given and choose Activate online. You can also import a .lic file. You do not send back a machine ID or any personal information. Linux has no activation window. macOS can activate in the window or from Terminal with ./StreetSegmentation activate YOUR-CODE (or activate --file code.lic). The server and renewal rules are the same.

The first activation needs the internet. After that the computer can work offline, but the app tries to renew about every 7 days and must come online at least once every 30 days. One activation code binds to one computer only. Copying the local licence file to another computer does not work, and colleagues cannot share a code.

If you change computers, replace the hard drive, or reinstall the operating system, the old code will not activate the new machine. Email 1252485024@qq.com and ask for a replacement; we sell a new code at half price when you change machines.

Single-image mode: five steps

• Import: On the Single Image tab, click Open Image or drag one image in. jpg, png, bmp, webp and tif are supported, including Chinese paths.
• Scene: Cityscapes for ordinary roads (19 classes). Mapillary for finer street detail such as lane lines, barriers and manholes (65 classes). ADE20K for indoor rooms, parks and mixed indoor–outdoor scenes (about 150 classes). Click Details for the full class list.
• Model: keep SegFormer B0 if you are unsure. Use B2 on a stronger machine, B5 only with a capable NVIDIA GPU (the app will confirm first). Mask2Former is slower and finer. Mapillary currently uses the official Mask2Former Swin-Large checkpoint only. Older CNN models such as FCN, PSP-Net, DeepLab v3+ and SegNet are not included — SegFormer and Mask2Former already outperform them on street-view work. Click Details for parameter count, weight size and hardware advice.
• Run: Click Segment Current Image. Image Reconstruction stays folded under Advanced Settings; Adaptive is the default. Very large photos are shrunk in memory, and if memory is still insufficient the app reduces the long edge and retries. The file on disk is never changed.
• Review and export: above the preview, switch Compare / Mask / Labels / Original, then drag mask opacity left for more of the photo, right for more colour; 100% shows pure colour blocks. Label maps sit on the original photo with class names; type size follows the image, and Label text can switch to English for journal figures. Export Current Result when ready.

Many images at once

Switch to the Batch tab, select an image folder, then choose the scene and model. Optional metrics and Advanced Settings (image-reconstruction mode) match the single-image tab. Click Start Batch Segmentation. This mode does not load per-image previews. Results are exported automatically to 分割结果/批量 beside the app (single-image exports go to 分割结果/单张). Nested input folders are preserved. Class shares and any ticked metrics go into one 分割结果.csv at the job root — not one CSV per photo. If the app directory is not writable, the Documents folder is used instead. When processing finishes, the app shows the full output path and an Open Results Folder shortcut.

Progress is checkpointed atomically after every successful image. Selecting the same input folder again resumes the existing job when the scene, model, opacity, output types and metrics all match; image-reconstruction settings are intentionally excluded from the configuration fingerprint. Completed images are skipped only when their path, size and modification time are unchanged and every expected output still exists. Another configuration change, edited source or missing output makes that image run again; rows in 分割结果.csv are updated rather than duplicated.

Batch processing supports Pause / Continue and Stop. Completed exports and checkpoints are retained after stopping. Single-image processing only provides Interrupt: no partial state is cached, so starting again processes the whole image from the beginning. A running CPU/GPU forward pass cannot be killed safely, so a control may take effect at the next safe checkpoint; the interrupted current image is not exported.

CLI guide

You can run the same installed app from Terminal without opening the window. You do not install Python. Subcommands are the same on macOS and Linux: activate, run, batch, and so on. On macOS, cd into 语义分割助手（街景版）.app/Contents/MacOS. On Linux, unpack the .tar.gz and cd into StreetSegmentation (keep _internal next to the program). Then run the commands below. Double-clicking the Mac app still opens the window. Linux has no window. Add --help on any subcommand for the full flag list.

./StreetSegmentation activate YOUR-CODE
./StreetSegmentation run photo.jpg --dataset cityscapes --model b0

What the exported files are

PNG files and the CSV footer include a licence ID so a result can be traced to an authorised copy. That ID is not an activation code and should not be shared as one.

Optional metrics are selected before segmentation and stay off by default. Colour richness and visual entropy use the original photo; green-view, enclosure, sky openness, building frontage, road share, vehicle share and walkable-ground share use the segmentation. Water-surface ratio needs Mapillary or ADE20K. Click “?” beside a metric for its formula.

• *_mask.png — colour-block map: solid class colours, no photo underneath.
• *_overlay.png — mask overlay: the photo with class colours on top, using the current opacity.
• *_label.png — label map: original photo plus colour, class boundaries and names. Type size follows the image; switch Label text to English for journal figures.
• *_id.png — class-id image for later software; skip it unless you need it.
• *_labels.csv — single-image export only: open in Excel to see pixel counts per class. Optional scene-metric columns are added only if you tick them.
• 分割结果.csv — batch jobs write one combined table at the job root, one row per photo, with size, ticked metrics and every class share.

Licence: one code, one computer

One activation code equals one computer. The code binds to the machine you first activate. Sharing it with a colleague, copying files to another PC, or pasting it after you change machines will fail.

Changing computers, replacing the drive, or reinstalling the OS looks like a new machine. Do not keep retrying the old code. Email 1252485024@qq.com, say you need a replacement for a new computer, and we will sell you a new code at half price.

The window shows a short licence summary, including when it must come online again. If you stay connected, renewal is usually automatic after about a week. If you see a message that the online lease has expired, connect to the internet and open the app or run segmentation once more. Do not post the code in a group chat.

Why not FCN, PSP-Net, DeepLab v3+ or SegNet

Those names show up in textbooks and older papers (roughly 2015–2018). They are omitted on purpose. On current street-view work, SegFormer and Mask2Former are ahead across the board: higher mIoU on Cityscapes and ADE20K, cleaner edges on thin or small objects such as lane lines, barriers and people, and official street-view weights that are actually maintained.

SegFormer B0 already runs on an ordinary laptop without a dedicated GPU. Shipping the older models would only enlarge the installer. If a past assignment used DeepLab v3+ or PSP-Net, start with SegFormer B0 here; switch to Mask2Former when you need finer boundaries. Do not look for the old names just to match an older paper title.